This PDF file contains the front matter associated with SPIE Proceedings Volume 13180, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Infrared images have the advantage of revealing thermal signatures, enabling enhanced visibility and detection of objects, especially in low-light conditions or scenarios where conventional cameras may struggle. However, detecting small moving objects, such as cars, drones, and people, in infrared images is a challenging task affected by ground clutter and complex image backgrounds, which increases more false positives in detection results. To relieve this problem, a twostage moving object detection method is proposed in this paper. Firstly, a newly designed attention mechanism is inserted into the feature extraction backbone for the YOLOv8 model, which helps the detection framework focus on the object region to reduce interference from complex image backgrounds and increase the detection rate. Secondly, an FP reduction strategy is proposed based on object moving analysis, which associates the detection results in the previous frame of the video with the current frame and performs target motion pattern analysis, and FPs that do not conform to the movement patterns are removed. Experimental results show that the proposed method is beneficial to not only reducing the FPs but also increasing the detection rate.

The existing image information hiding schemes have the shortcomings of insufficient embedding capacity and weak robustness, which are the key factors restricting their development. In order to address this problem, this paper proposes an image information hiding scheme based on interleaving permutation. In the proposed scheme, the secret image is processed by using interleaving permutation firstly; and then the secret image after interleaving permutation are compressed by using differential pulse code modulation (DPCM). Finally, the compressed bits are embedded into the low bit-plane of the carrier image. Experimental results show that the proposed scheme significantly expands the embedding capacity in comparison with the existing image information hiding schemes.

Existing super-resolution reconstruction algorithms for remote sensing images often struggle to fully extract and utilize features in complex scenes, and the reconstruction results are not optimal due to the influence of noise. We propose a reconstruction network model that combines deep dense residual module and sub-pixel convolution. This model connects multiple dense residual modules through recursive linking and introduces a channel attention mechanism to extract multiscale features from the images. During the feature reconstruction process, a sub-pixel convolution structure is introduced to reduce the impact of noise and enhance reconstruction performance. The model is tested on the UC Merced Land Use public dataset, and the results demonstrate that the reconstruction results of our proposed model algorithm have better peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) at different scales compared to the current mainstream reconstruction algorithm EDSR. It is proven that the proposed model can significantly improve the reconstruction quality through comparative analysis experiment, meeting the needs for high-resolution remote sensing image processing.

Computer vision-based methods can detect fatigue through eye movement and mouth state analysis. Due to the accuracy, real-time, non-invasive and low cost of image vision methods, the method of recognizing specific facial images has been proved to be the most industrialized technology at present. However, when people are tired driving, factors such as external light, head posture and facial changes still bring many challenges to human fatigue detection. The most important part of fatigue driving detection is face detection, feature extraction and classification recognition. This paper mainly studies the key problems of extracting facial features based on CNN network and infering fatigue state in real vehicle operating environment by LSTM network. The experimental results show that the proposed method has high performance and can be further applied to practical scenarios.

With the rapid development of the civilian drone market, the phenomenon of illegal drone flights is escalating, making regulation increasingly challenging. To safeguard public safety, the tracking and countermeasure technology for drones has become an urgent priority. However, due to various constraints, obtaining specific drone images in complex backgrounds is challenging, especially under different lighting conditions. Additionally, the number of available images is insufficient to meet the training and validation needs of tracking algorithms and other requirements. In this context, this paper, inspired by the attention mechanism, proposes a visible light UAV image simulation method based on Generative Adversarial Networks (GANs). This method transforms physical simulation images generated by computers into drone images with authentic styles under different lighting conditions. Subsequently, the Mask R-CNN image segmentation model is employed to accurately separate the generated drone targets from the original background, and the Poisson blending technique is used to seamlessly merge them with different backgrounds. Experimental results indicate that our approach generates drone images with features closer to real images across various evaluation criteria. Furthermore, after merging with different backgrounds, the images exhibit high coherence and visual realism, providing a viable solution for effectively expanding the datasets of drone tracking algorithms.

The fusion of millimeter-wave radar and optical camera is a challenge in the multi-sensor data fusion field. In this paper a new space calibration algorithm is proposed to match the low-resolution radar point cloud information and the optical image information. This paper proposes that fusing the wheeled odometer and radar to increase the density of point cloud. Moreover, the real 3D model information of marker and radar point cloud are fitted to extract the 3D feature point of marker, and the 2D pixel feature point of the marker is extracted by Instance Segmentation algorithm. Then, the marker’s 3D-2D matching points are used to solve the external parameters between the radar and camera. Based on Turtlebot2 hardware platform, the experimental results show that the proposed method can calibrate the radar and camera, and match the radar point cloud and optical image information under the condition of large resolution difference. The proposed algorithm has practical application value in autonomous driving, such as target detection and tracking, scene 3D reconstruction, path planning and navigation.

Aiming at the problem that existing low-illumination image enhancement methods cannot simultaneously take into account multiple degrading factors, which leads to poor processing effect when applied to real scenes, deep neural network models facing different image processing tasks are integrated based on neural network integration strategy, and multi-task and multi-stage cascade processing is performed. The low illumination image can be enhanced under the interference of multiple degrading factors. The experimental results show that the proposed method can significantly improve the overall brightness and contrast of the image, and the enhanced image has natural colour distribution and good visual effect, which is obviously superior to the existing mainstream low-illumination image enhancement methods.

Synthetic aperture radar (SAR) is an essential tool for ocean surveillance. As the main participants on the ocean, ships are the most important targets for ocean monitoring. So it is of great importance to develop ship detection algorithms for SAR sea images. The algorithms based on Convolutional Neural Network perform far better than the traditional methods based on manual features on ship detection task due to the powerful feature representation abilities. The algorithms based on Convolutional Neural Network can be divided into one-stage algorithms, and two-stage algorithms. Two-stage algorithms have high accuracy, but are relatively time-consuming. One-stage algorithms have high inference speed, but compared with two-stage algorithms, they have lower accuracy. So in this article, we proposed an modified one-stage detection algorithm to improve the accuracy of ship detection in a condition that the modified algorithm meet the real-time requirement. First, the small model of one-stage algorithm YOLOV5 is chosen as the base network to get the high inference speed. Then, to improve the accuracy of ship detection with a little increase in inference time and the model parameters, we integrate the one–layer super-resolution architecture with the simplist structure into the YOLOV5 network. Finally, we conducted the comparative experiments on our Dataset to verify the performance of modified YOLO V5. The experimental results show that the modified method has obtained an Average Precision (AP) improvement than the original YOLO V5 for detecting ships in SAR images with a little increase in inference time and the model parameters.

Addressing the challenges inherent in remote sensing image detection, notably the YOLOv5 detector's subpar accuracy due to limited detection target features, intricate detection backgrounds, and predominantly small targets, this paper presents the RS-YOLOv5 algorithm, an enhancement of the YOLOv5s model. Primarily, the conventional convolution in the backbone network is substituted with full-dimensional dynamic convolution to enrich the information gathered across multiple dimensions, thereby enhancing the extraction of target features. Additionally, a decoupled detection head replaces the traditional detection head, refining classification and positioning tasks while strategically leveraging features from different levels to mitigate background interference on small target detection. Next, adding the Normalized Wasserstein distance (NWD) measurement to the loss function is intended to enhance the accuracy of small target detection. The ablation analysis conducted on the DOTA dataset showcases a noteworthy improvement in the average accuracy of object detection in remote sensing imagery facilitated by the proposed RS-YOLOv5 algorithm, culminating in a 2.5% increase in average accuracy compared to the original algorithm.

This study suggests a multi-scale polarization fusion method based on lifting wavelet transform to enhance the clarity and contrast of underwater polarization images, which are currently low contrast, of poor quality, and blurred texture features. In the experiment, underwater polarization imaging of several materials using a 532 nm blue-green laser was used, and intensity photographs taken from various angles were collected. The polarization degree and polarization Angle of the image are calculated, and the polarization combination image is generated according to the method of region energy maximization. The polarization combination image and light intensity image are decomposed into low frequency components reflecting contour information and high frequency components reflecting detail information at multiple scales. The regional energy as the weight coefficient and the fusion rules based on the combination of region variance are constructed and applied to the low frequency and high frequency components respectively. Experimental simulations demonstrate that this fusion method has some advantages over conventional image fusion techniques in terms of boosting image contrast and clarity. It has also established a strong base for resource discovery, bottom topography exploration, submarine detection, and other sectors.

Referring image segmentation aims to segment an object mentioned in natural language from an image. It is a fundamental computer vision task. This task is challenging because it involves both vision and language features that need to be aligned and fused effectively. For alignment, pre-trained CLIP is widely used in many vision-language tasks for its notable success in aligning these two modalities. However, in the majority of existing methods, vision and language information are independent in the encoder stage, which is a suboptimal fusion approach. In this paper, we introduce an innovative CLIPDriven Hierarchical Fusion framework named CHRIS. We utilize CLIP as the encoder for its valuable vision-language alignment, we also design an effective early fusion approach in the encoder stage called hierarchical attention. Moreover, we introduce a novel hierarchical fusion neck to fuse vision and language information. In this way, the vision and language features contained in CLIP are further fused effectively. We perform comprehensive experiments on the three datasets widely adopted in the research community, RefCOCO, RefCOCO+, and G-Ref. Our proposed framework demonstrates superior performance compared to previous approaches by just using ResNet as the backbone.

Diffuse optical tomography (DOT) guided by medical images can be used to achieve real-time reconstruction of subdural hematoma images for monitoring purposes. However, when the hematoma is irregular and the signal to noise ratio(SNR) of input signal is low, the reconstruction effect of this method is not ideal. In order to alleviate this situation and improve the reconstruction effect of cerebral hematoma images, we proposed a optical image mapping feature extraction method in the paper. The result of experiment shows that the mean relative volume error (VRE) of the hematoma reconstruction model optimized by the optical image mapping feature extraction method is only 0.79%, and the mean value of the average Manhattan distance (AMD) between the reconstructed absorption coefficient and the true absorption coefficient is 0.0062mm-1. Compared with the model directly inputting optical information, the optical image mapping feature extraction method reduces the average VRE of the model by 93.3%, and the average AMD by 27.1%. This method provides a promising method for non-invasive continuous monitoring of clinical cerebral hematoma.

Comprehending the intricacies of image structure and noise distribution is pivotal in employing paired training methods. Nevertheless, these approaches encounter difficulties when attempting to generalize to images characterized by unknown noise distributions. In response to this challenge, we propose a two-stage image denoising method designed to enhance generalization performance. In the initial stage, we introduce a preliminary denoiser based on a Multilayer Perceptron (MLP). This denoiser utilizes an implicit structural prior to forecast preliminary denoising results, ensuring alignment with the clean image structure. This stage necessitates training with a limited set of paired low-noise images. In the subsequent stage, we incorporate initial denoising results as guiding conditions for denoising diffusion null-space model (DDNM), we proficiently harness the generated diffusion prior, markedly amplifying the denoising model's generalization capability. We utilize a pretrained unconditional diffusion model, obviating the need for supplementary training or network optimization, thereby resulting in an economical overall training cost for the entire method. Through extensive experimental validation encompassing diverse datasets, noise types, and intensity levels, our method consistently outperforms alternative image denoising techniques in both denoising performance and generalization ability.

As an extension of diffusion models, consistency models reduce the necessary sampling steps to a single iteration when synthesizing an image sample, thereby significantly enhances the efficiency of image generation. In addition, they also allow multi-step generation, providing flexibility for trade-offs between sample quality and computational efficiency. Despite these advantages, traditional consistency models are troubled by loss of high-frequency image details. This issue is attributed to the inherent regression-to-mean property of the L2 training loss, impeding overall improvement of model performance on the image generation task. In this paper, we propose a novel consistency adversarial model to address the loss of high-frequency image details through adversarial generation. In particular, we train a consistency model in an adversarial manner by treating it as an image generator. Then, an additional image discriminator is introduced and optimized along with the consistency model in an adversarial manner. The target of the image discriminator is to punish the image generator when it synthesizes images lacking high-frequency image details. In this way, image samples with high-frequency image details can be obtained and the performance of consistency models can be improved. Extensive experiments demonstrate the effectiveness of our proposed method. Our CAM outperforms the traditional consistency model on two challenging benchmarks: ImageNet and LSUN.

Automatic segmentation of temporomandibular joint (TMJ) images helps physicians to analyze the joint structures and is a critical step in the diagnosis of temporomandibular joint osteoarthrosis (TMJ-OA). However, the boundary of the TMJ condylar usually cannot be accurately segmented, since the bone density of the condyle is low, which means the condyle approximates the background in Cone Beam CT (CBCT) images. In this work, we proposed Double-level Boundary Enhanced Network (DBE-Net) by adding the Double-level Boundary Enhanced (DBE) module to the U-Net network to enhance the boundary segmentation accuracy of condylar joint boundaries in CBCT images by providing additional guidance. The DBE module that connects multi-stage features from the encoder of U-Net, which can simultaneously emphasize lower-level features with fine-grained spatial details and higher-level features with rich semantic information thus enabling the module to focus on boundary information processing. Experimental results on the clinical dataset comprising 88 subjects with TMJ CBCT images, our method has achieved Dice of 0.9761 and the 95% Hausdorff distance of 0.2483.

Multi-temporal collaborative analysis of port scenes can enhance the representation ability of image scenes, and image registration is required before multi-temporal analysis. In this paper, an image registration network E-SuperGlue with enhanced feature matching is proposed to solve the problems such as the difficulty of extracting feature points and matching feature descriptors for port multi-temporal image registration. Our network takes SuperGlue network as the basic framework. Firstly, Focus is introduced into the feature extraction network to increase the number to increase the number and detection rate of feature points. Secondly, LFPE module is added to feature matching network coding module to improve the information efficiency of feature descriptor coding. Finally, an improved multi-layer sensing structure E-MLP is added to the feature matching network to improve the utilization rate of channel information.

Because the current method is not ideal in digital landscape mapping, the mapping error is high, and the expected mapping effect can not be achieved. Therefore, the application of three-dimensional laser scanning and oblique photography technology in digital landscape mapping is proposed. Firstly, the distribution characteristics of landscape are collected, and the average value of sample characteristics is calculated, then the rules of landscape construction are output and the results of landscape distribution characteristics are calculated. The UAV is equipped with a three-dimensional laser scanner to update the relative altitude, simplify and compress the point cloud data, measure the transverse scanning angle and longitudinal scanning angle of each laser pulse signal, and obtain the three-dimensional coordinates of the measured point; Analyze the overlapping situation of oblique photographic images, build a garden digital mapping model, calculate the external orientation elements of the images, and complete the garden digital mapping. The experimental results show that the error of surveying and mapping data obtained by this method is small, and the matching degree between digital landscape mapping and actual landscape mapping site is high.

This paper propose a retinal image segmentation model according to hybrid pooling and multi-scale attention mechanism. Due to the crucial importance of retinal vessels for the segmentation of the retina and their intricate structure, which serves as vital diagnostic information for diseases, we focus on retinal images in medical imaging. We introduce a novel medical image segmentation model, MDA-Unet, built upon the Unet architecture. Extensive experiments are conducted on the DRIVE dataset. Addressing challenges such as limited data, varying image quality, and the presence of small structures in segmented target areas causing reduced segmentation accuracy, the proposed enhanced model MDA-Unet aims to overcome these issues. We investigate the impact of incorporating modules such as Exponential Moving Average (EMA), Multi-Path Module (MPM), and serpentine convolutional kernels on the model. The research results indicate that the addition of these modules enhances the ability of the model to be segmented. This improved model is tested with the DRIVE dataset, yielding segmentation results that surpass those of previously proposed segmentation models. Compared with the previous model segmentation results, the method proposed in this paper has achieved good results in fundus vascular segmentation.

Sperm detection in forensic evidence has become crucial in proving sexual assault. The current method in Thailand relies on manual microscopic examination by experts to visually identify sperm cells from specimens. This process, while effective, is time-consuming and susceptible to inaccuracies due to artifacts resembling spermatozoa. Therefore, this research explored the benefits of incorporating standard forensic laboratory methods to automate and expedite sperm detection. This research utilized image processing for automated and rapid sperm identification, thereby assisting medical experts in making fast and efficient diagnoses. It presents a classification algorithm based on color, texture and morphology, employing computer vision and image processing techniques including the Hough Transform. The algorithm achieved an accuracy of 95.00%, a recall of 94.12%, and an average efficiency of 96.97% compared to manual examinations by forensic experts. Moreover, it efficiently processed images at an average speed of 0.199 seconds per image.

Image super-resolution reconstruction (ISRR) can reconstruct a high-resolution image from low-resolution images, and ISRR algorithms based deep learning (DL) networks appear to be one of the most promising directions. Motivated by recent studies, this paper summarizes the ISRR algorithms based on DL, with a focus on convolutional neural network algorithms, deep residual network algorithms and other typical algorithms. Based on the above analysis, a very deep super-resolution (VDSR) algorithm is proposed, and experimental verification on the DIV2K dataset is conducted as well. Furthermore, the important highlights and prospects of DL-based ISRR algorithm innovation are introduced.

Feature point detection algorithms have been widely used in the fields of object recognition, panorama stitching, and robot navigation. SIFT algorithm is a robust feature detection method widely used in image processing and computer vision. This design proposes a SIFT algorithm based on FPGA development platform implementation, which reduces the computational complexity and improves the processing speed by adopting a pipelined architecture. The matching results show that the algorithm has good invariance to image rotation, illumination, affine, scale and can meet the needs of feature matching, and there is a certain practical application value.

Remote sensing images are widely used in various fields, such as geographic information systems, environmental monitoring, and agricultural management. However, remote sensing images are often corrupted by various noises during acquisition, transmission, and storage, which degrade the image quality and affect the image analysis and utilization. Therefore, de-noising remote sensing images is an important step to improve the image quality and application performance. In this paper, we propose a remote sensing image de-noising method based on wavelet packet decomposition and information entropy threshold function. The proposed method uses wavelet packet decomposition to perform multiresolution analysis on remote sensing images, and combines information entropy threshold function to adaptively suppress noises. The experimental results show that the proposed method can effectively remove Gaussian noise and salt-andpepper noise, and preserve the edge and detail information of remote sensing images. The proposed method outperforms the existing methods in terms of peak signal-to-noise ratio, structural similarity index, and visual quality.

The utilization of long-range spatial features has significantly enhanced hyperspectral image (HSI) classification performance. However, capturing global spatial relationships among pixels comes with a high computational expense. Given that super-pixels are created adaptively from the content within hyperspectral images, leveraging them for global spatial information extraction has proven to be an efficient strategy. Nevertheless, some methodologies at the super-pixel level, like Graph Convolutional Networks (GCN), are burdened by numerous parameters and substantial computational complexity. In order to address these challenges, this paper introduces an Efficient MLP-Aided CNN Network with Localglobal Feature Fusion for Hyperspectral Image Classification (EMACN). EMACN adopts a dual-branch structure, utilizing a super-pixel-based Multilayer Perceptron (MLP) for global feature extraction with a pixel-based Convolutional Neural Network (CNN) for local feature extraction, thereby achieving an effective balance between computational efficiency and classification performance. Compared to the quadratic complexity seen in other super-pixel level deep learning structures such as GCNs, the computational complexity of our super-pixel-based MLP approach is linear with respect to both the number and dimensionality of super-pixels. Furthermore, we enhance our model by integrating a spectral transformation subnet that utilizes group convolution instead of the traditional 1x1 convolution for reducing the dimensionality of hyperspectral images. Additionally, our model incorporates a pixel-based CNN module that employs 2D depthwise separable convolutions, optimizing computational efficiency. This design allows our network to achieve an optimal balance between its representation capabilities and computational expenses. Extensive testing across three datasets confirms the superior performance of our proposed network.

To improve the application efficiency of RGB remote sensing images in agricultural land resource surveys, a cultivated land segmentation algorithm based on kernel space non-uniform regularization classification and improved graph cut was proposed. Firstly, extracting texture and color features of remote sensing images using Local Binary Pattern algorithm (LBP), Gabor filters, and RGB, HSV color space, respectively. Next, introducing a kernel method to map data from lowdimension to high-dimension, and construct a kernel space-based non-uniform regularization sparse representation model to classify and segment images in pixel level. Finally, an innovative graph cut algorithm is enhanced by incorporating a Gaussian distribution to redefine the penalty term for homogeneous regions and introducing a new color gradient measure to define the penalty term for boundaries. This approach effectively removes scatter and restricts the segmentation boundary. The average classification accuracy and average F1 score of the classifier proposed in this paper are about 2% and 3% higher than those of recent regularized subspace classifiers, respectively. Compared with the Graph cut algorithm, the proposed improved graph cut algorithm has an average mIoU improvement of about 9%. The average accuracy of the whole segmentation algorithm is 95.43%, and the average mIoU is 88.56%. Compared with the comparison algorithm, the proposed algorithm has higher segmentation accuracy, which proves that the proposed algorithm can adapt to the cultivated land segmentation scene of remote sensing images and is effective.

Intertidal terrain extraction based on satellite remote sensing images is of great military application and civilian value. In this paper, we propose a fusion algorithm based on SAR and optical remote sensing image to enhance intertidal feature. With the SAR and panchromatic images from different phases, we use an improved shift-invariant discrete wavelet transform algorithm to obtain higher quality images that can not only highlight the shape of the intertidal zone, but also highlight the characteristics of the coastline. The real data experiment and comparison experiments with other algorithms show that the algorithm proposed in this paper is effective in enhancing the features of the intertidal zone.

We introduce a deep learning methodology that combines the Inception-ResNet-V2 architecture with a hybrid attention mechanism. This fusion enables the extraction of crucial information from both channel and spatial dimensions within feature maps, giving rise to the ECA-SA-Inception-ResNet model. To assess the performance of our model, we conducted experiments on a petrographic thin section image dataset utilizing Grad-CAM visualization techniques. The results demonstrate the model's proficiency in extracting essential features from petrographic thin section images, thereby resulting in substantial enhancements in image classification accuracy.

In recent years, methods based on deep convolutional neural networks (CNNs) have gradually become the focus of research in the field of hyperspectral image (HSI) classification. It is well known that hyperspectral data itself contains spatial and spectral information. While CNN-based methods have advantages in extracting local spatial features, they are not good at handling spectral features and global information. Therefore, this paper proposes a multi-attention network that fuses local and key channel information to complete the task of HSI classification. First, the principal component analysis (PCA) is used to pre-process the HSI data. Second, a feature information fusion module based on the SE module and 2D convolution is constructed to fuse local spatial information and enhanced feature channel information. Third, the global covariance pooling function accelerates the convergence rate of the network. Finally, the fused features are sent to the Vision Transformer (ViT) module for position encoding to capture global sequential information and improve the hyperspectral image classification results. Experiments carried out on several typical three public datasets demonstrate that the proposed network method can provide competitive results compared to the other state-of-the-art HSI networks.

Currently, reconstruction-based visual anomaly detection tends to use more complex models to improve the accuracy of the model. This paper proposes a lightweight image restoration network to reduce the network model's parameters and computational complexity, and puts forward the feature similarity loss to reduce the impact of noise on the accuracy of anomaly detection discrimination. Experimental results show that the proposed method uses fewer parameters and computational complexity, achieving results close to RIAD.

In today's hospitals, doctors often need to use medical equipment to find the cause of the disease, and the clarity of medical images often affects the doctor's judgment of the patient's condition. Having high-resolution medical images can help doctors better treat patients. But nowadays, most super-resolution models cannot effectively restore medical images, resulting in poor image restoration results. This article follows the idea of cyclic generative networks and trains them using Wasserstein distance to solve the task of image reconstruction. The model consists of an image reconstruction network, an image degradation network, and two discriminators. In the generative network, we implement cyclic consistency based on Wasserstein distance to establish a nonlinear end-to-end mapping from noisy LR input images to denoised and deblurred HR output images. This article experimented and validated the performance on the FastMRI dataset, and compared with existing mainstream methods, the results showed that it was superior to current mainstream methods in medical image super-resolution reconstruction.

Due to the increasing scale of the reference landmark image database, the efficiency of image retrieval tasks is significantly reduced. In this work, we propose an image retrieval method based on feature clustering to improve the efficiency of largescale image retrieval. Firstly, The Deep Orthogonal Local and Global(DOLG)information fusion framework is used to extract features from images. Secondly, the optimal number of clusters is determined by Bayesian. Then, the image features are divided into different clusters by the Gaussian mixture clustering algorithm. After that, the feature of the queried image will be compared with all centers of clusters and retrieved in the luster with the nearest cluster center. Finally, the retrieved images are reordered by the contextual similarity aggregated with the self-attention mechanism; the most similar referenced images are regarded as the retrieval result. Experimental results on the Roxford5k and Rparis6k datasets demonstrate that the proposed feature clustering-based image retrieval method can significantly reduce the time cost by more than 26% without loss of accuracy.

At present, with the in-depth exploration of image super-resolution reconstruction technology promoted by deep learning, the model structure is becoming more and more complex, which leads to the reduction of the feature learning ability of the model, and the number of model parameters and calculation amount are greatly increased. Therefore, this paper proposes a combination of channel and spatial attention mechanism, residual feature fusion module, and improved knowledge distillation algorithm to enhance feature learning ability and reduce the number of redundant parameters in the network. The experimental results show that this method can effectively strengthen the feature learning ability of the model and reconstruct the high-frequency details, and realize the parameter compression of the model, and reduce the parameter calculation.

The extraction of building locations is crucial in the field of remote sensing, commonly applied in tasks such as emergency response, urban planning, and environmental monitoring. Existing methods often employ convolutional models such as ResNet and U-Net. however, these models struggle to capture long-range features of geographical objects, limiting their practical effectiveness. Meanwhile, Transformer based models face challenges due to their quadratic computational complexity. Additionally, the characteristics of remote sensing images, including dense arrangement, small scale, and class imbalance, pose difficulties for traditional patch-merging modules in downsampling, leading to the loss of significant building information. To address such limitations, we present the HSDFormer model in this paper. The model achieves a reduction in the computational complexity of self-attention calculation through a straightforward sequential reduction process. Additionally, our model has optimized the downsampling process specifically for remote sensing semantic segmentation tasks, effectively minimizing the loss of crucial information during layer-wise feature extraction. We evaluate the HSDFormer on the WHU Building Dataset and Satellite Dataset II (East Asia). The results demonstrate the model’s effectiveness in accurately segmenting building locations, surpassing performance compared to other baseline models.

A polarization image frame capture device based on Camera Link interface is proposed and implemented. The polarization image frame capture device adopts large-capacity buffer device, multi-bus switching and DSP technology, which solves the problem of multi-channel asynchronous image parallel acquisition and high-speed real-time polarization image synthesis. It can capture and synthesize the larger polarization image frames in real time.

The traditional F5 steganography algorithm hides the information in the whole image and uses matrix coding for information embedding, which can embed a large amount of information with little change, and has high embedding efficiency, but lacks robustness to compressed JPEG images. In this paper, a new F5 steganography technique is proposed, which combines F5 steganography with the combination of lossless carrier and complexity coefficient to process JPEG images. First, this paper constructs a lossless carrier to enhance the robustness of F5 steganography, then measures the complexity of the image by the complexity coefficient, and selects the complex pixel blocks for F5 steganography to further enhance the robustness of JPEG image steganography.

Image fusion is a significant research area, related to a specific fusion task and has broad application prospects. Most of the existing image fusion algorithms are based on the pixel level, and although the convolutional network with more layers has a powerful feature extraction capability, its complexity increases with the deepening of the network. The dependency of the local range of the image is also not fully utilized in this process by the convolutional network, which makes the fused image detail information lost. In this paper, we propose a local-attention mechanism based network with multi-scale residuals to fuse visible and infrared images. The network consists of two key parts: encoder-decoder, fusion strategy. During the network training process, we utilize a phased training approach, where an automatic codec is first trained for conducting feature extraction, local feature enhancement and feature reconstruction. In the fusion stage, the coder trained in the first step is utilized to extract the two light image features, and then these features are fused by a multi-scale residual network, respectively. Finally, the fused features are inverse encoding to acquire the fused image. The experimental comparison results show that our fusion solution outperforms the existing fusion methods in both visual perception and objective evaluation.

Following the rise of deep learning, neural networks are increasingly used in the field of image rain removal. However, most of the rain removal models proposed so far have problems of poor rain removal quality and low efficiency. Based on PReNet, this paper conducts improvement research from the aspects of lightweight and multi-scale feature extraction, proposes the multi-scale progressive recurrent network MsPRN. The main work and contributions are as follows: Use a progressive cyclic network architecture to significantly reduce the amount of parameters and achieve lightweight models through recursive loops and parameter reuse; Propose the Multi-cascade Progressive Atrous Convolution module (McPAC), which uses atrous convolution and multi-cascade channel separation to lightweightly extract deep feature information and reduce rain streak residue; It is proposed to use the squeeze-and-excitation (SE) module to selectively enhance or suppress the propagation of feature channel information, so as to accurately restore the image background details. Experimental results on the Rain100H, Rain100L and Rain12 synthetic data sets show that MsPRN can achieve better rain removal effects and performance while keeping the model lightweight.

With the continuous improvement on the refined management of coal-fired power plant, it is a necessary task to accurately obtain the total quantity of coal in power plant. In this paper, ORB (Oriented Fast and Rotated Brief) - edge coupling method was proposed for the coal pile to strengthen the weak texture boundary of coal pile images, and thus improve the feature extraction and matching capacity. The results showed that our proposed algorithm can increase the number of feature points by a factor of about 10 times relatively to the ORB algorithm. Furthermore, three types of coal piles were examined to evaluate the number of feature points. The results showed that the accuracy rate of feature point matching can reach up to 90%, indicating that our proposed algorithm has the good effectiveness and applicability.

For detecting prohibited items inside packages, visual image doesn’t work well as items overlapped or occluded. Thus, Penetrable X-ray image was employed to solve the occlusion problem by coloring items in different material, yet the overlapping problem still remains unresolved. Therefore, this paper proposes an X-ray image segmentation and metal prohibited items detection following a coarse-to-fine scheme. Firstly, the overlapping metal objects are clipped according to the value in the H channel and R channel. Secondly, individual metal items are extracted using regional growing. Thirdly, metal prohibited items are detected by matching its shape with those predefined in Database. Our experiment show that the proposed method can detect 94.74% metal prohibited items inside packages.

The aim of this paper is to investigate cervical cancer image processing technology utilizing deep learning. Cervical cancer stands as a prevalent malignancy in females, and precise identification and localization of cancer cells hold paramount significance for treatment and prognosis evaluation. This paper presents the fundamental workflow of cervical cancer image processing and the associated principles of deep learning, including convolutional neural networks, autoencoders, and generative adversarial networks. In recent times, the swift advancement of deep learning technology has brought forth novel concepts and approaches for cervical cancer image processing. This paper is oriented toward the exploration of cervical cancer image processing technology grounded in deep learning. First, the basic workflow of cervical cancer image processing, including steps such as image acquisition, preprocessing, feature extraction, and target detection, is introduced. The application of deep learning in cervical cancer image processing is discussed in detail. As one of the core deep learning technologies, convolutional neural networks (CNNs) have achieved significant results in the fields of image classification, segmentation, and detection. This paper shall present the fundamental principles and prevalent architectures of CNNs, alongside their instances of utilization in cervical cancer image processing. Furthermore, the utilization of alternative deep learning approaches in cervical cancer image processing is also introduced. Subsequently, the paper contrasts the strengths and weaknesses of diverse deep learning techniques in cervical cancer image processing and deliberates the challenges and future trajectories of development within this domain.

In order to realize the accurate recognition of landslides in remote sensing images, an improved DeepLabv3+ landslide extraction model is proposed in this paper. (1) Hybrid Module and Attention Module based CSPNet (HA-CSPNet) is constructed as the backbone feature extraction network to enhance the feature information of small landslides and suppress the interference of irrelevant background information. (2) Combining the advantages of Residual Structure and Dense Connected Module, Residual-dense ASPP is designed to focus landslide features at different scales, enhance feature reuse and prevent gradient vanishing. The experimental results show that the landslide extraction model proposed in this paper is practical. It can improve the accuracy of landslide semantic segmentation.

Automated breast ultrasound image segmentation enhances radiologists' ability to enhance breast cancer diagnosis accuracy. In view of the inherent speckle artifacts, fuzzy breast lesion boundaries, and inhomogeneity within breast lesions in ultrasound images, this paper proposes a multiscale residual UHRNet network based on a global multiscale attention mechanism (GMS-UHRNet) to enhance the segmentation accuracy of breast lesions in ultrasound images. On the basic framework of UHRNet, the feature expression ability is effectively improved by parallel multiscale residual fusion approach, which enhances the segmentation accuracy and robustness of the network. In the GMS-UHRNet network, a global multiscale attention mechanism is added, which successfully reduces the semantic difference with the encoder and enables the network to recognize breast lesion regions more accurately. Extensive experiments on the BUSI dataset demonstrate the network's superior performance in breast ultrasound lesion segmentation compared to other medical image segmentation methods.

Radiology report summarization is a core part of the computer-aided-diagnosis, which aims to automatically generate examination impressions to accurately summarize key medical information in basic radiology reports. Most existing studies improve the accuracy of report summarization by deeply mining internal connections of basic radiology reports. However, these methods simply regard radiology report summarization as a simple text-to-text single-modality task, ignoring the indispensable role of raw medical images in the entire report summary work. To address this issue, we propose a novel radiology report summarization framework based on joint embedding of graph-guided features. Our method uses a Topic Graph to accurately capture disease information in medical images, and a Word Graph to deeply mine important relationships between keywords in radiology reports. By using Graph Convolutional Networks (GCN) to model the key information of the two modalities, we successfully achieve effective fusion of the two features and ultimately generate highly accurate report impressions. Experimental results on two benchmark datasets OPEN-I and MIMIC-CXR demonstrate the effectiveness of our method. Our method demonstrates superior performance over existing works on multiple evaluation metrics.

The illicit trade of prohibit products poses significant health and economic challenges globally, prompting the need for more effective detection methods in express shipment inspections. This study introduces a novel multi-dimensional approach to prohibit package detection by combining X-ray scanning imagery with simulated express shipment information, leveraging the capabilities of both computer vision and large language models (LLMs). We employ a state-of-the-art feature extraction model, specifically a You Only Look Once (YOLO) variant, to analyse self-constructed X-ray image datasets for contraband prohibit indicators. Concurrently, we generate a simulated express information dataset, encapsulating patterns characteristic of prohibit smuggling tactics. This data is then integrated with image-derived features through custom-designed prompts to train LLM classifiers. Our methodology is unique in its consideration of multi-modal data fusion, hypothesizing that the synthesis of visual and textual information will markedly improve detection accuracy over traditional single-dimension analysis. The experimental results demonstrate the efficacy of this approach, with the LLM classifiers outperforming standard methods in accurately identifying prohibit packages. The study not only provides a new perspective on the application of LLMs in contraband detection but also sets a precedent for future research in multimodal data exploitation for security and customs enforcement.

Because of the convenience and low cost of obtaining visible remote sensing images of UAV, it is widely used in agricultural production. In land cover classification, in order to obtain more homogeneous superpixels of UAV remote sensing images and achieve accurate classification, a superpixel segmentation algorithm with improved DPC (density peak clustering) of grid point was proposed, and land cover classification was carried out based on this algorithm. The method is as follows: the paper defined color-space distance measurement to calculate local density and relative distance; then, using the connected graph, the connected distance between each grid point to be allocated and the cluster center point connected to the neighborhood is obtained, and it is assigned to the superpixel cluster with the smallest distance. Finally, pixel blocks are generated according to the super-pixel segmentation image and as data to classification by resnet50. Experiments show that the classification accuracy of the proposed algorithm is 99.6%, which is superior to SLIC and the original superpixel segmentation algorithm with DPC of grid point, and proves that the algorithm is effective.

Surface defect detection plays a crucial role in the quality management of precision industrial products and holds significant research value. Defect detection in industrial environments still needs to address two main challenges: first, the proportion of product defects on the surface is too small to capture, presenting a typical small target detection problem; second, some products have complex surfaces with numerous defects, making it difficult to monitor each defect individually. Aiming to address the aforementioned issues, this paper proposes a solution for surface defect detection based on image processing. It successfully detects 11 types of defects on object surfaces and applies this method in industrial settings. First, a lighting platform was built, and defect pictures were collected using industrial cameras and integrated into picture datasets. Secondly, the dataset is preprocessed using image processing to identify four types of defects, and then further processed through image enhancement. The graph is then sent to the deep learning network for training, and the model parameters are adjusted based on the defect state to achieve a convergent model. The experimental results demonstrate that the proposed method can accurately identify 11 types of defects on the product surface with a precision of 98.5%, satisfying the requirements of industrial settings.

This paper introduces a groundbreaking Neural Architecture Search (NAS) technique, addressing the critical balance between computational efficiency and model performance within the context of Convolutional Neural Networks (CNNs). Recognizing the significant impact of architecture on CNN performance, and the computational intensity of evolutionary NAS methods despite their superior global search capabilities, we propose a novel solution. Our method integrates two key innovations: an Early-Quit (EQ) strategy to reduce the model's Floating Point Operations Per Second (FLOPS) and the incorporation of channel attention mechanisms into the NAS search space. Unlike traditional approaches that embed these mechanisms directly into the network, our strategy minimizes computational waste by terminating training early for less promising architectures and enhances feature extraction through attention mechanisms. Applied to standard benchmarks like CIFAR-10 and CIFAR-100 our approach demonstrates a substantial decrease in FLOPS while maintaining or even improving learning efficacy. This NAS methodology represents a significant advancement by efficiently generating lean yet competitive neural network architectures, marrying computational efficiency with strategic innovation.

Thanks to the feature disentanglement capability of disentangled representation methods, deep learning-based de-entanglement methods have achieved strong performance in image fusion tasks. However, it should be noted that training a powerful performance of a disentangled representation method is a challenging problem due to the domain differences between image fusion tasks and mainstream disentanglement tasks. To overcome this problem, an infrared and visible image fusion method based on cycle consistency and the covariance compression disentanglement module proposed in this manuscript. Firstly, the proposed covariance compression disentanglement module statistically counts the global information by global covariance pool and adaptively enhances the desired features and suppresses others by the covariance compression module. Secondly, it makes the disentanglement module cycle-consistent by disentanglement the fusion image, and strengthens the disentanglement by the proposed consistency loss function and feature similarity loss function constraints. Specifically, in this paper, multi-source images are input into the disentanglement module, modal features and common features are disentangled, and fusion results are generated by fusing the modal features and common features of multi-source images. After generating the fusion result, the fusion result is input into the disentanglement module again, and the re-decomposed features are used to calculate the consistency loss with the multi-source image features. On the publicly available TNO dataset, subjective and objective evaluations show that the method proposed in this manuscript has better fusion performance than other state-of-the-art methods.

Convolutional neural networks (CNNs) are the mainstream model for extracting rich features in deep learning-driven studies on cloud detection for remote sensing images. However, due to the limitation of receptive fields in convolutional operations, a single CNN has limitations in mining global information and long-range dependencies, which affects the precision of cloud identification in intricate environments. To address the insufficient global feature and long-range dependency modeling of CNNs, a new CNN-Transformer network for cloud detection (CNN-TFCD) in remote sensing images is proposed in this paper. This method comprehensively utilizes strengths of both CNN and transformer to fully mine local spatial details, global contextual information, and long-range dependencies, to obtain comprehensive feature representations. The feature cascade mechanism is applied for multi-scale feature fusion, further enhancing the model's feature representation capability. Evaluations conducted on the Landsat-8 dataset demonstrate that the CNN-TFCD method achieves accurate and stable cloud detection.

Image description refers to the process of automatically generating natural language descriptions that are strongly related to the content of an image, and it is a cross-disciplinary field that combines computer vision and natural language processing. This paper proposes an improved attention mechanism for image feature fusion, which addresses the limitations of existing image feature extraction methods. The proposed method uses an encoder-decoder structure, where an improved GAM attention module is used to fuse the grid features of images with edge features. Furthermore, a mesh memory structure is employed to further enhance the fused features, resulting in richer image features. Through the decoder, more accurate image descriptions can be generated. The model was evaluated using mainstream evaluation metrics BLEU, ROUGEL, CIDEr, and METEOR, and validated on the public MS COCO dataset. Experimental outcomes demonstrate that the image description model proposed based on GAM, which merges different features, achieves favorable performance across various evaluation criteria and further enhances the capability of image description.

Esophageal cancer is a common digestive tract tumor with a high mortality rate. Currently, the identification of esophageal carcinoma predominantly hinges on thoracic computed tomography scanning. Because the area of the esophagus on CT images is small and belongs to the small target region, the traditional image segmentation network is difficult to accurately segment the tumor area of esophageal cancer. Although the traditional U-Net architecture and Transformer integrated variants perform well in medical segmentation tasks, they lack image localization and channel feature extraction capabilities. To address the aforementioned issues, this manuscript proposes a novel medical image segmentation model, christened DAA-TransUNet. We propose a new module called Dual Attention Area (DAA) module,it can extract image position and channel features. And we combine the DAA module and the Efficient Channel Attention (ECA) module into a new module, which we call the Efficient Channel Dual Attention (ECDA) module. DAA-TransUNet integrates Transformer and ECDA modules into a traditional U-shaped architecture, enabling it to extract not only global and local information, but also image position and channel features. We have verified the superior performance of our network by performing numerous experiments on our collected esophageal dataset. We used DAA-TransUNet to perform experiments on our collected dataset, and the average DSC(%) reached 75.26% and the average HD95(mm) reached 3.54.

Medical images often consist of multiple modalities, such as multimodal MRI images commonly used in diagnosing and studying brain tumors., and multimodal images provide rich complementary information. In the past, multimodal image segmentation usually directly added or connected modal features in the early or middle stage, which made it difficult to obtain the connection between modal features. In addition, there is a difference in information between modals and modals, and the previous method did not dealign modal features, which is likely to lead to reduced the effect of modal fusion. Thus, we propose a Multiscale dual dynamic feature fusion transformer (MdcFormer) model to explore the effects of multi-scale features, spatial and channel dynamic fusion and modal feature alignment on the segmentation effect of multimodal medical images. Utilizing a multi-encoder configuration and a single decoder, we gather characteristics from various modes at various levels and blend them in a dynamic manner across both spatial and channel domains. The proposed approach was evaluated using the BraTS2020 benchmark dataset. Empirical findings indicate that the model enhances the precision of segmentation.

When images are captured in dark environments, they may be of low quality due to extreme lighting conditions and equipment limitations. The degradation of underexposed images includes high noise in dark areas, low visibility, and loss of details. Many enhancement methods have been proposed, but generally suffer from inconsistent color and incomplete detail, with artifacts appearing in extremely dark areas. In addition, these methods are computationally expensive due to the use of hierarchical processing methods. Therefore, an integrated, noise-sensitive approach is needed. This paper introduces the Invertible Enhancement Network (InvEn-Net), which uses the Invertible network and BM3D algorithms to separate the high-frequency and low-frequency components. The compositional relationship between the two components is learned based on the original image. The low-frequency components undergo color correction and partial detail supplementation using a U-shaped convolutional neural network (UNet). The prior distribution (Gaussian distribution) of the high-frequency components is resampled, finally leading to the restoration of the normally exposed image. Compared with various types of low light image enhancement methods, our framework can achieve competitive color consistency and structural similarity (SSIM). Ablation experiments justify our selection of model structural parameters.

Infrared images are widely used in security monitoring and autonomous driving due to their resistance to light changes and adverse weather conditions. But, the low contrast and colorless characteristics of infrared images limit the effectiveness of human observation and subsequent detection and recognition algorithms. Methods for translating infrared images into visible images can overcome the above shortcomings, among which contrastive learning methods using self-similarity features achieve the best performance. However, methods based on contrastive learning face the sub-optimization problem caused by random sampling. And the contrastive loss based on self-similarity features faces the problem of encoding entanglement when used for infrared-visible image translation, that is, the features extracted from different categories of regions cannot be distinguished. Therefore, we propose a cross-similarity guided contrastive learning method for infrared-visible image translation. First, to address the randomness and inefficiency of the contrastive loss random sampling process, a sampling strategy based on information entropy ranking of cross-similarity matrix is proposed to obtain sampling points for subsequent contrastive loss calculation. By calculating the information entropy of cross-similarity matrix between input and generated images and sorting them, the sampling points with the most information can be obtained. Second, to alleviate the encoding entanglement problem of the self-similarity contrastive loss due to the low contrast of infrared images, multi-scale spatially adjacent graph structure consistency loss and spatially separated graph structure consistency loss based on cross-similarity matrices are proposed. Experiments on KAIST and FLIR datasets show that the proposed method has the best score and visual performance compared with multiple advanced infrared-visible translation methods. Ablation experiments further illustrate the effectiveness of the method.

Deep learning (DL) techniques have been widely applied in medical image analysis. In particular, the DL-based medical image classification has been adequately investigated on large-size annotated datasets. However, it is cost-expensive to collect a large amount of high-quality and large-scale annotated medical images. Our proposal is addressing this problem by a few-shot medical image classification method that uses contrastive learning and linear discriminant analysis (FSCCLLDA). A well-performing encoder is pre-trained using contrastive learning to extract more extensive semantic information that is unrelated to the label. Moreover, the features are transformed into low-dimensional space using linear discriminant analysis (LDA). The transformed features are similar within each class and discriminatory among classes. Experiments on ISIC2018 and BreakHis datasets show that the proposed FSC-CLLDA algorithm outperforms the compared baselines in accuracy.

3D reconstruction is greatly important for perceptual understanding of applications in robotics, computer graphics, and human-computer interaction. Though tremendous researches have been carried out for geometry reconstruction, there are still many challenges to cope with the large-scale, featureless, and specular surfaces. In this paper, we are motivated to propose a unified framework of geometry reconstruction and fusion for surfaces with the above characteristics. For the geometry reconstruction of specular surfaces, we present a method combining Stereo Phase Measuring Deflectometry (SPMD) with modal wavefront reconstruction to retrieve a reliable geometric surface. For the geometry fusion of largescale featureless surfaces, we put forward an offline calibration approach exploiting the Structure-from-Motion (SfM) to merge point clouds reconstructed from different viewpoints. To evaluate the effectiveness of our method, we build a specific SPMD camera system, design a large-scale calibration board, and set up a test scene with an industrial robot and car body. The experiments show that our method can achieve a good performance in terms of reconstruction accuracy of around 0.02mm and fusion accuracy of around 2mm for a large surface.

The substantial scale variation and intra-class diversity within remote sensing imagery pose significant challenges for semantic segmentation, rendering methods developed for natural images inapplicable. These challenges, we introduce a novel semantic segmentation model named D-CANet, which primarily comprises three modules: the Global Class Center Awareness (GCCA), the Local Class Awareness Module (LCAM), and the Global Class Generation Module (GCG). Specifically, the GCCA module is dedicated to modeling the global representation of class context to mitigate the interference from image backgrounds; the LCAM module generates a local class representation, serving as an intermediary perceptual element that facilitates an implicit linkage between pixels and global class representations, minimizing the variance within classes; following the processing by the LCAM module, the GCG module enhances the global class representation. This encoder-decoder structure equipped with GCCA, LCAM, and GCG modules achieves precise segmentation of objects of varying scales within remote sensing imagery through the interactive perception and fusion of global and local features. Experimental assessments conducted on the Potsdam dataset and the Vaihingen dataset illustrate that D-CANet surpasses the current state-of-the-art semantic segmentation techniques in terms of efficacy.

Uncooled long-wave infrared detectors suffer from temperature drift and image quality degradation due to the lackof aconstant temperature environment. A shutter is commonly used to calibrate the detector periodically, but it causes wearand tear on the lens and detector, and may fail in harsh environments. This article presents a more advanced researchmethod for nonuniformity correction based on the blackbody image space in uncooled long-wave infrared cameraswithout a shutter. The proposed method integrates based on calibration and based on real time scenario non-uniformitycorrection techniques in two step: offline calibration stage builds an blackbody image space fromblackbody imagescollected at different ambient temperatures; online identification stage finds the optimal projection of the scene imageinblackbody image space using the statistical prior of the sharp real time scenario infrared image. And estimates theoptimal blackbody image for non-uniformity correction at the current ambient temperature. It creates a “virtual shutter”in front of the lens that effectively compensates for the temperature drift of the detector characteristics andlens. Comparative experiments show that the proposed method significantly reduces non-uniformity and adapts to the ambient temperature variation.

The Antarctic ice sheet is an important part of the Earth system and plays a key role in the global climate system. The frequency modulated continuous wave (FMCW) radar can be used to obtain the changes of the transmitted signals at different layers inside the ice sheet, and then the information of the isochronous layers of the ice sheet can be extracted and the shallow accumulation rate can be deduced. In this paper, an improved spectrum estimation algorithm is proposed, which combines FFT (Fast Fourier Transform) and Chirp-Z (CZT transform, linear frequency modulation transform) to estimate the spectrum of the simulated radar system's difference frequency signals, effectively improving the radar detection accuracy, and verifying the effectiveness and rationality of the algorithm, so that the isochronous layers can be extracted on a larger spatial scale. In China's 36th Antarctic scientific expedition, FMCW radar was used to obtain the distribution of shallow isochronous layers in East Antarctica from Zhongshan Station to Taishan Station. In this paper, taking a typical area as an example, the algorithm was used to obtain the stratification and structure information of the ice sheet on a large-scale space, which provides important data support for the calculation of large-scale accumulation rate and flow history in the Antarctic region.

Fine-grained classification poses significant challenges due to high intra-class variability and high inter-class similarity, making coarse-grained classification methods inadequate. Previous research has focused on locating objects in images while overlooking their detailed features. Additionally, images contain not only targets but also redundant information such as background and noise. The relationships between extracted features are also crucial, a point often overlooked in previous studies. To address these challenges, this paper proposes the Multi-scale Graph Neural Network Filter (MGF) model, composed of two modules: the Filter High-Resolution Feature Pyramid Network (FHRFPN) and the Adaptive Graph Neural Network (AGNN). FHRFPN progressively fuses features of different scales to retain rich detail information while avoiding conflicts between information and locating target positions. Before outputting, it employs the Channel Filter (CF) block to filter out noise features and reduce unnecessary computational overhead. AGNN adaptively creates adjacency matrices between features based on the input features. Through graph neural networks, the model comprehensively considers the interactions between features, effectively capturing non-linear and complex dependencies. The proposed method achieves state-of-the-art performance on the NA-Birds and CUB-200-2011 benchmarks, thus providing a promising solution for enhancing the performance of fine-grained visual classification tasks.

Emotion recognition from physiological signals is a crucial area in affective computing. However, traditional CNN models face challenges in accuracy and efficiency. This paper proposes a lightweight IGC-CNN model that integrates interleaved group convolutions with the LeNet-5 network. Experimental results using EEG, EMG, and EDA signals collected across happiness, sadness, and fear states show that IGC-CNN achieves an average accuracy of 94.74%, outperforming traditional CNNs by 10.06%. Statistical analysis confirms the significance of this improvement (P < 0.01). Evaluation metrics such as AUC, precision, recall, and F1 score further validate the superior performance of IGC-CNN. This study suggests that IGC-CNN is a promising approach for multimodal physiological signal-based emotion recognition.

When ray tracing is performed based on the traditional KD-Tree (K-Dimensional tree) structure, a ray may first pass through a number of leaf nodes before it intersects with one facet in the “last” leaf node. During this process, the ray needs to sequentially perform the ray-triangle intersection with all the internal facets in each of the aforementioned leaf nodes before it enters into the next leaf node, wasting a large amount of computation time. In this paper, an improved KD-Tree structure is proposed. In addition to the space occupied by a leaf node in the traditional KD-Tree, the minimum bounding box occupied by the facets of the leaf node is also introduced for the leaf node. If a ray does not intersect with the minimum bounding box when it enters into a leaf node, it will directly penetrate out of the leaf node without performing ray-triangle intersection procedure and then enter into the next leaf node. Simulation results for T-cavity and L-cavity models show that, under the parameter conditions of this paper, the total ray tracing time based on the improved KD-Tree structure is reduced by 10% to 44% compared with the traditional KD-Tree, which effectively improves the computational efficiency of the ray tracing process.

Due to the complexity, chaotic behavior, and non-linear nature of tokamak plasma dynamics, modeling tokamak discharges poses a formidable challenge. This modeling task necessitates the accurate prediction of the evolution of tokamak diagnostic signals, entailing the resolution of a sequence-to-sequence time series prediction problem. In this study, we propose a novel framework for tokamak discharge modeling, termed the PCC-Trans framework, which integrates a time series feature selection method based on the Pearson correlation coefficient (PCC) and Transformer, and enhances the predictive accuracy of the model. Specifically, PCC is employed for signals correlation analysis and the selection of relevant, non-redundant features to augment the Transformer. Through analysis and modeling of experimental data from the Experimental Advanced Superconducting Tokamak (EAST), the proposed framework demonstrates superior performance across four case studies compared to five other machine learning approaches, including Multi-layer Perceptron (MLP), Long Short-Term Memory (LSTM), Bi-directional Long Short-Term Memory (BiLSTM), Convolutional Neural Network (CNN), and Transformer. Additionally, we validate the efficacy of the feature selection method by assessing the necessity of each element within the feature set and its influence on modeling performance.

Unmanned Aerial vehicles (UAVs) are increasingly used in healthcare, agriculture, environment, disaster management and other fields. Therefore, it is very important to research and build a system that can accurately identify UAV signals in a complex environment. This paper aims to explore a UAV signal recognition system based on the classification of the signals sent by the UAV and its controller. Wavelet transform is used to extract features, and neural network encoder is used to classify UAV signals and controller signals to realize the distinction between UAV signals and controller signals, different UAV models and different controller models. This model adopts the architecture of an autoencoder and introduces innovative design in two frameworks. The initial framework incorporates skip connections within the residual layers, ensuring the preservation of original information during convolutional operations. Meanwhile, the second framework capitalizes on this preserved information to extract supplementary features. These designs collectively enable the model to achieve a more profound understanding of the feature information embedded within the data.

Aiming at the problem of poor classification performance caused by the loss of useful information in the process of feature extraction of moving image brain computer interface (MI-BCI). This study describes IHCL-Net, an improved hybrid convolutional neural network (CNN) and long short-term memory (LSTM) network. We show that IHCL-Net effectively reduces feature loss during model training and accurately extracts spatiotemporal features from EEG data. Our proposed network accurately maps the temporal correlation of signals to a new space during the extraction of spatial features, which preserves the integrity of temporal characteristics as much as possible. Preliminary results showed that the IHCL-Net had the average accuracy value of 85.3% on IV-2a compared with other networks, achieving the highest classification performance.

Visual segmentation based on the meaning of Uyghur text is an important process of dynamic Uyghur character retrieval， It also provides great convenience for workers who learn and understand the Uygur language for the first time. To address the challenges of strong continuity and high character similarity in handwritten Uyghur text, which make effective segmentation difficult, we propose an efficient and lightweight model for handwritten Uyghur text Detection segmentation, based on the Deeplabv3+ network in convolutional neural networks. Our experiment selects 100 common words in Uyghur language, and collects a total of 7093 handwritten samples. We use median filtering, adaptive threshold binarization and other operations to standardize and denoise the samples to achieve scientific preprocessing. We use the MobileNetV2 module by replacing the original Xception module in the model, to guarantee precision, significantly reducing training parameters and saving training time. To solve potential positive and negative sample imbalance in data, we use Focal loss and Dice loss. Moreover, we introduce CBAM attention mechanism in the deep feature extraction network to enhance the ability of deep feature extraction and refine important text features. Our experimental results demonstrate that the proposed algorithm significantly improves segmentation accuracy while meeting performance requirements, shortening training time, and enhancing the segmentation effect of handwritten Uyghur text, improve readability of split text.

Multi-scale exposure fusion is an effective way to directly fuse low dynamic range (LDR) image with different exposures into a content-rich LDR image for high dynamic range (HDR) reconstruction. Previous researches have shown that edge-preserving smoothing can be used to improve multi-scale exposure fusion. However, multi-scale exposure fusion via edge-preserving smoothing pyramids suffers from loss of details. To address this issue, we propose a side window gradient guided image filtering (SGGIF) and use it to construct an edge-preserving smooth pyramid. First, by adding eight kernels to the gradient guided image filtering(GGIF), a SGGIF with effective edge preserving is developed. Furthermore, we select the weight map with the minimum mean as the guidance image, which can further preserve details in the brightest and darkest regions of HDR scenes. Finally, we developed a detail-preserving multi-scale exposure fusion method based on edge-preserving smooth pyramids. Experimental results indicate that our method can effectively preserve details and reduce halo artifacts. Both quantitative and qualitative analyses demonstrate the effectiveness of our proposed approach.

To facilitate self-examination for businesses and enable government agencies to monitor food safety based on sentiment analysis of takeaway reviews, this paper proposes an RoBERTa-WWM-based sentiment analysis model for the takeaway industry, termed RLMT. The model converts the text expressions in takeaway reviews into dynamic semantic feature representations using the RoBERTa-WWM model. The study integrates the global semantic features extracted by combining BiLSTM (Bidirectional Long Short-Term Memory Network) with the Multi-head Attention mechanism and the local semantic features derived from TextCNN (Text Convolutional Network). The fused features are then processed through a fully connected layer and the Softmax function to obtain the final sentiment polarity. Experimental results illustrate that the RLMT model demonstrates superior sentiment classification performance on the waimai_10k dataset compared to current popular sentiment analysis methods, with precision, recall, and F1 values reaching 93.62%, 92.62%, and 92.62%, respectively.

Cross project defect prediction is a valuable strategy for identifying software defects when historical data is unavailable, yet selecting the right training project for the predictive model is a challenging task. Additionally, dealing with numerous redundant features during model training significantly impacts prediction accuracy. To tackle these challenges, this paper presents a solution that combines candidate project selection with feature filtering. It combines the Collaborative Filtering method with the Content Based method to identify the most appropriate candidate project. For feature filtering, the Pearson feature selection approach is utilized to eliminate data duplication and determine the optimal number of features. Experimental results affirm the effectiveness of this approach, showcasing its superiority to existing approaches in regard to F-score and Area Under the Curve metrics.

FPGA is an emerging hardware productivity that is characterized by high flexibility, reconfigurability and moderate cost. The development of FPGA EDA software is relatively lagging behind due to the problems of difficult programming and excessive compilation time, which constrains the upward expansion of the FPGA field. In this paper, we aim to solve the problem of long compilation time of FPGA EDA software, and make an in-depth study on routing, which is one of the most time-consuming steps in the compilation process. A combination of coarse- and fine-grained FPGA parallel routing algorithm is proposed. Coarse-grained routing uses a load-balanced space partitioning strategy to generate a routing sequence tree, and when routing each net mesh, it switches to a fine-grained parallel algorithm to perform parallel routing within the net mesh with one source-sink node pair. Experiments show that the CFPara algorithm proposed in this paper compared to the VPR PathFinder algorithm can obtain a 11.7× speedup ratio.

Optical remote sensing cloud detection is the basis of quantitative remote sensing and remote sensing application, Aiming at the problems of single cloud detection method and cumbersome label making, The multi-channel data of the geostationary orbit radiation Imager (AGRI) on the domestic FY4A satellite was selected as the research object, Cross-polarized cloud-aerosol Lidar (CALIOP) was used as the reference object, Achieve all-weather, all-terrain and efficient cloud detection, Through spatio-temporal matching, the overall accuracy of FY4A official cloud detection product is about 85.1%. Through regional similarity, image processing and other methods, Improved EfficientNetV2 model. The overall accuracy of the model is more than 89.9%, It is higher than the accuracy rate of FY4A official cloud detection products, providing a new reference for the research of cloud detection related products.

Video-text retrieval has been a crucial task with the exponential growth of video. Recent methods that leverage the pretrained CLIP model in the video-text retrieval task have demonstrated remarkable performance, surpassing many approaches trained on large-scale video-text datasets. However, these works ignore the performance-efficiency trade-off in the pursuit of better performance. Additionally, a completely new model is typically necessary for each task because many existing works fully fine-tune the pre-trained backbone. Therefore, to yield a compact and transferable model, we propose LoCLIP, a framework that transfers knowledge from CLIP in a parameter-efficient manner. Inspired by LoRA, we incorporate only a small set of trainable low-rank matrices per task, allowing adaptation to new tasks by simply replacing these matrices. In this way, we can acquire task-specific knowledge without compromising the prior knowledge stored in the pre-trained backbone. To demonstrate the effectiveness of our LoCLIP, we conduct extensive experiments and achieve comparable performance with state-of-the-art CLIP-based video-text retrieval methods while updating only a few parameters.

This paper proposes an improved TransUNet medical image segmentation method for tongue positioning and segmentation to solve the premise of tongue image diagnosis task and the segmentation performance of deep learning network. By introducing efficient multi-scale attention module in the jump connection layer of the network, the image conducts deeper feature extraction and compensate for the spatial information lost by image dimension reduction. At the same time, an improved empty space convolution pooling pyramid is added to the convolutional neural layer of the network to enhance the network perception of the input image while reducing the network training cost. Facing the complex background conditions and a variety of interference problems, the network has excellent performance.

Due to the scarcity of point cloud datasets in a specific domain, utilizing generative model approaches becomes essential for data augmentation. Diffusion models have demonstrated impressive capabilities in data generation through a guided reverse process. In this work, we employ a reverse process of a Markov chain conditioned on shape latent to progressively generate dental crown point cloud from a noise distribution. We propose to map the global shape latent to a set of partlevel implicit representations and introduce a cross-attention block to provide geometric structural information for point cloud generation. We conduct a series of experiments on a real dental crown dataset, and the experimental results show certain improvement compared to the baselines, demonstrating the efficacy of our approach. In experiments, we present the capability of our method to generate large-scale dental crown models through unsupervised learning, effectively enriching the existing dental crown dataset.

The accurate prediction of low-visibility has important application value to the normal operation of navigation, aviation and other transportation vehicles. However, the current low-visibility prediction task mostly uses time series model to extract temporal features, but neglects the spatial correlation between national observation stations. Additionally, those time series model require that the input data must be limited to the Euclidean space. However, due to the discrete distribution of stations, the input data required for multi-station low-visibility prediction does not have a regular spatial structure. To address those limitation, we proposed a weighted location feature convolution based on graph convolution. The construction of the weighted location feature convolution takes into account the distance and pitch angle, reduces the influence caused by the dispersion of national observation stations, and enables the convolution kernel to be shared in different local meteorological spaces. The convolution as a substitute for matrix multiplication within Sequence to Sequence framework, then cascaded the decoder to deal with the spatio-temporal dependence relationship of different levels, leading spatio-temporal fusion model CPGDC. The experimental results show that the CPGDC model achieves has higher accuracy in predicting low-visibility values across multiple national observation stations.

The question generation task is a type of task in the NLP field that uses the answer of the question as the input for a model, allowing the model to trace back to the corresponding original question. Prompts can be designed to guide the large language model to generate questions that are closer to its corresponding answers. Previous work involves adding textual content to the prompt to clarify specific and help the model understand, however, there is still room for improvement in the quality of the questions generated by these prompts. On the basis of previous work, this paper adopts the method of automatic question generation and includes few-shot examples as part of the model input. We automatically generated 20 basic prompts and proposed three methods for constructing demonstrations:(1) Most helpful shots (2) Maximum entropy shots (3) Mixed method. We used GPT-3.5-turbo as the experimental model and selected two suitable datasets TruthfulQA and Alpaca-cleaned for this task. In the stage of evaluating the quality of generated questions, we used text cosine similarity and GPT-4 scoring to evaluate the impact of designing demonstrations on the quality of generated questions before and after. We compared the generated questions of the model with the preset questions corresponding to the answers in the dataset. Quantify the difference using cosine similarity: The improved result is 3.09% higher than the original prompt's generation question result ; we used the GPT-4 large language model for intelligent scoring, which also showed an improvement in scores on both datasets. The study also explored the impact of different numbers of shots under a specific method of the task. When the number of shots is 6, the combined result of text similarity on two datasets was 0.828, achieving the best result. When the number of shots is 5, the best GPT-4 rating result is 4.435.

Affected by the multiplicity of the sum and rotation of mixed music signals, the corresponding reliability is difficult to be guaranteed when recognizing their frequencies. Therefore, a study on mixed music signals and rotation frequency recognition based on improved S transform is proposed. On the basis of comprehensive analysis of the composition of mixed music signals and spins, the mixed music signals and spins are processed by time-frequency conversion using Stransform. In the phase of frequency recognition, a cyclic neural network (RNN) model is constructed, and the input dimension of the model is set to be consistent with the dimension of extracted and spins frequency characteristics, and the output dimension is consistent with the number of spins, The mixed music signal and rotation can realize frequency recognition by forward and backward propagation in the model. In the test results, the recognition result accuracy rate is stable at more than 77.5%, and the recall rate is stable at more than 76.0%. Compared with the control group, it has obvious advantages.

This study proposes a GrabCut algorithm to address the segmentation challenges in coffee leaf images. The algorithm integrates the Contrast Limited Adaptive Histogram Equalization algorithm and the Gamma Correction technique. The leaf images undergo preprocessing using Contrast Limited Adaptive Histogram Equalization and Gamma Correction before being segmented with the GrabCut algorithm. Experimental evaluations are conducted using coffee leaf images, comparing results in terms of intuitive segmentation visualization and quantized indices such as PSNR, segmentation accuracy, false segmentation rate, and algorithm runtime. The findings indicate that the proposed algorithm exhibits greater robustness against variations in light conditions, resulting in higher segmentation accuracy. It outperforms classical three-class benchmark algorithms and effectively segments leaf images.

Remote sensing technology has become one of the most effective means of acquiring land cover information, benefiting from its capability to provide dynamic, rich, and cost-effective data sources. Additionally, the changes in vegetation coverage are crucial for forest management, ecological conservation, and government planning. Thus, this study utilizes remote sensing image data to analyze the vegetation cover changes in Puyang, central China. During the period from 2002 to 2017, three scenes from the series Landsat, i.e. Enhanced Thematic Mapper Plus (ETM+) and Operational Land Imager (OLI), are used in the experiment. After that, the Normalized Difference Vegetation Index (NDVI) values are derived from these data. An empirical model is employed to classify the study area into five different vegetation cover levels, i.e., no vegetation, extremely low coverage, low coverage, moderate coverage, and high coverage. The proportions of each category in the study area are calculated in three satellite images, respectively. The experimental results indicate that the vegetation coverage of Puyang is low, emphasizing the need for further strengthening of environmental protection. Specifically, during the period from 2002 to 2009, there was a noticeable decline in vegetation coverage in Puyang, accompanied by a substantial increase in bare soil and built-up land areas. Then, in the subsequent period from 2009 to 2017, there was a significant rise in the proportion of areas with low vegetation coverage, indicating the tangible effectiveness of environmental protection measures. The analysis suggests that these changes are closely related to the economic transformation in Puyang. Hence, to better safeguard the ecological environment of Puyang, it is recommended that future environmental protection strategies place greater emphasis on enhancing vegetation coverage and promoting sustainable development.

Poyang Lake is the largest freshwater lake in China. It is located in the middle and lower reaches of the Yangtze River in the north-central Jiangxi Province. As the largest Lake which is connecting the Yangtze River, the water ecological security of Poyang Lake affects the entire Yangtze River basin. Water transparency(SDD) is an important indicator of water quality, so it is of great significance to carry out routine monitoring of SDD using remote sensing methods. This study takes Poyang Lake as the experimental area, and uses the SDD as the inversion index to characterize water quality. Based on observation data from 2015 to 2020 and Landsat series images, the inversion of SDD model suitable for summer Poyang Lake was constructed by analyzing and comparing the correlation coefficients between different wave bands and the SDD under the condition of standardized modeling phase and "star-ground" collaborative matching. After verification by four sets of independent data, the mean R2 is 0.8621, the mean RSME is 15.37cm, and the mean MAPE is 23.6%. It is believed that the method used in this study can be well applied to the inversion of SDD in the Poyang Lake area, and also provides a new reference method for monitoring the transparency of inland water bodies under long-term observation conditions.

In the acoustic monitoring of belt conveyor rollers, multiple sound sources and strong noise often impede remote fault diagnosis based on acoustic signals. To address this issue, this study proposes an underdetermined blind source separation method that combines singular spectrum decomposition (SSD) and improved sparse component analysis (SCA), which is suitable for tackling noise interference in roller acoustic inspection systems. Field tests have validated the effectiveness of this method, which reduces reliance on extensive a prior knowledge and the need for numerous sensors, demonstrating outstanding performance in separating sound sources of faulty rollers from multiple noise sources.

The article proposes a method for bearing fault diagnosis based on the waveform FFT spectrum of the dual-stream CNN. The method first uses FFT and wavelet transform to obtain the one-dimensional FFT spectrum and two-dimensional time-frequency map of bearing vibration, and then inputs them into the 1D-CNN channel and 2D-CNN channel of the model for feature extraction. After the feature information is fused and processed, the fault diagnosis is finally completed in the classification layer. The experimental results show that this model significantly improves the accuracy of bearing fault diagnosis.

Summary Object detection is an important direction of computer vision, and small Object detection has long been a difficult point in computer vision. With the progress of satellite and remote sensing technology, there are a large number of highdefinition remote sensing images every day. Because the target's feature map will shift after the target passes through the deep convolution network, which has a great impact on small target detection. In order to solve this problem, we introduce the receptive field module RFB in the relevant algorithm to enhance the discrimination of small target features. We further introduced the ASPP spatial pyramid module into the backbone network to expand the receptive field of the backbone network, so as to strengthen the detection of small and medium-sized targets. In order to evaluate its effectiveness, we carried out experiments in YOLOX target detection algorithm. The results show that this improvement can significantly improve the detection accuracy of small targets without affecting the reasoning speed.

The paper examines the integration of AI in visual communication design, particularly focusing on the enhanced impact of images when combined with text. It explores the evolution of visual communication design in the context of AI advancements, emphasizing the increasing significance of this discipline in today's technology-driven society. Key areas of discussion include the role of artificial intelligence, specifically deep learning, in image recognition and natural language processing, and its application in visual communication design. The paper details the development and training of an image recognition and text matching model, showcasing the effective application of convolutional neural networks (CNNs) and Word2Vec in visual design tasks. The research concludes by highlighting the successful translation of deep learning theoretical achievements into practical visual communication design applications, demonstrating the model's proficiency in image recognition and text matching within a business and engineering framework.

This paper presents a method to enhance the training of low-order grammar language models in response to the issue of out-of-vocabulary (OOV) words and incorrect words in acoustic model and low-order grammar language model decoding, which results in a decrease in Chinese speech recognition accuracy. To address the problem of incorrect words caused by polyphonic Chinese character, multiple phonemes are added to the Chinese dictionary to improve the phoneme model training. Prior to training the low-order grammar language model, the one-dimensional word segmentation is converted into a vectorized segmentation structure for processing. This enables the training of language models to obtain different probability distributions for the same word, thereby reducing the occurrence of OOV words. Finally, this method is compared with traditional language model training methods and evaluated with the same acoustic parameters. The experiment demonstrates an 8.5% reduction in word error rate for speech recognition when using low-order grammar language models.

There is a significant difference in correlated color temperature between deep-space environmental light sources and that on the ground. The lack of accurate calibration parameters for cameras can lead to errors in the collected scientific image data, resulting in the failure of scientific tasks. Thus it is of vital significance to accurate white balance calibration for deep-space-exploration cameras. However, traditional white-balance methods have certain limitations. The generalization ability of the white balance correction matrix is insufficient, the ground calibration methods have poor on-orbit adaptability, and deep space exploration targets are unknown and uncertain. To this end, we proposed an on-orbit white balance calibration method for deep-space-exploration cameras based on correlated correlated color temperature interpolation. Aiming to the deep space environment background, four pre-calibrated illuminations with different correlated color temperatures are selected reasonably. Before solving the white balance calibration matrix, we stretch the image grayscale in different exposure times to ensure the brightness balance and obtain the optimal white balance matrix, which improves the generalization ability of the matrix. Finally, by interpolating the pre-calibration white balance matrixes based on the on-orbit environmental correlated color temperature, the average color difference can be reduced by 28 (72.54%) DE2000 color difference units.

Few-shot learning (FSL) aims to learn new visual categories from a limited number of labeled examples. The use of mutual information is beneficial for improving the discrimination of feature embeddings, which makes metric classification more accurate. In this context, we propose Dense Masked Deep InfoMax Networks (DMDIN) based on Masked Image Modeling (MIM) framework with convolutional neural networks (CNN). DMDIN effectively incorporates the mutual information maximization strategy into self-supervised pretraining, improving its performance in few-shot learning. More specifically, DMDIN first generates two augmented views of each image through spatial transformations and color jittering for one view, and random RGB mean patching for the other. Second, DMDIN maximizes mutual information between dense features from these views, DMDIN learns more generalized representations, diverging from traditional patch reconstruction-based methods. Our method is evaluated on the miniImageNet, CIFAR-FS, and CUB datasets. The results validate the effectiveness of the proposed method in the few-shot learning scenario.

The test and inspection business of metrology field and laboratory in the electric power field faces the problems of low degree of automation of experimental data flow, high degree of heterogeneity of sensor transmission signals, low degree of resource management sharing and low degree of intelligence. In the paper, the CTC algorithm is used for report recognition, and deep learning is used to achieve the layout analysis of the report in different formats, and to complete the multi-mode data metrology testing image recognition processing and the construction of the knowledge base. The experimental results show that the image recognition technology proposed in the article can effectively extract the data in metrology inspection images and form a knowledge base.

With the acceleration of urban construction and renovation, the volume of construction waste has significantly increased, accounting for 30% to 40% of the total urban solid waste emissions. Therefore, analyzing the spectral characteristics of typical construction waste is crucial for remote sensing identification and monitoring, providing essential ground verification information. This study conducted spectral data collection of construction waste in areas such as Daxing District of Beijing and Pingdingshan City, Henan. Various spectral analysis methods were applied to examine its characteristics. The findings revealed that: (1) the wavelength range is between 1913nm and 1932nm, the spectral curves of all kinds of construction wastes have obvious reflection troughs, In the range of 1660-1690nm, organic building materials exhibit absorption valleys. (2) In the wavelength intervals of 600~1100nm and 1860~1930nm, spectral curves exhibit consistent trends, rendering them unsuitable for discrimination. (3) In the range of 1900~2200nm, samples exhibit reduced susceptibility to the spectral characteristics of the dust-proof net, closely mirroring the spectral features of uncovered samples, thereby rendering it suitable as a basis for classification.

In the fields of computer vision and natural language processing, cross-modal retrieval is of great importance that cannot be ignored. In existing multi-granularity alignment methods, significant progress has been made by globally aligning images and sentences or locally aligning regions and words. However, most of these methods rely on attention mechanisms, which may be affected by noise and cannot balance the differences between modalities, resulting in suboptimal image-text alignment. In addition, training with attention mechanisms consumes a lot of computational resources, and the retrieval process is time-consuming. To address these challenges, we develop a novel multigranularity image-text alignment model, which we call DFVLM. First, we independently train the intra-modal encoder and cross-modal encoder to reduce interference between different modal encoders and better learn intra-modal and intermodal information. Then, we propose a joint strategy to combine the intra-modal and inter-modal information to better capture the information between image and text. In addition, we introduce a hard negative pair-based method to train the performance of the intra-modal multi-granularity encoder without using attention mechanisms. Unlike existing methods, We treat image and text retrieval as a two-way process and delve into the intrinsic connections between the two matrices. We propose a cross-validation method to optimize the retrieval result based on common image-to-text alignment scenarios in daily life, i.e., during the image-to-text retrieval process, a text-to-image reverse verification is performed on the retrieved corresponding text. Through extensive qualitative experiments and analysis, our approach performs well on both Flickr30K and MSCOCO datasets and also significantly reduces the time required for testing.

Recent advancements in spectral reconstruction algorithms have yielded impressive results for RGB image restoration in hyperspectral tasks, utilizing Convolutional Neural Networks and Transformer networks. However, there are certain limitations associated with these approaches. The design of CNN networks is gradually developing towards deeper networks, which can hinder their ability to capture long-distance dependencies and similarities. Transformer networks effectively address this issue but introduce information redundancy, complex operations, and slow reasoning speeds. To address these concerns and strike a balance between information efficiency, algorithm accuracy, and real-time inference on end device, we propose a Hyperspace Reconstruction Unet (HsRu) network. Remarkably, the HsRu network features a minimal parameter count of only 200KB, constituting just 8% of the parameters found in the leading model, which has 2.25M parameters, yet it manages to maintain the root mean square error at a low 0.03. Unlike traditional Unet architectures, HsRu employs only two downsampling and upsampling stages to preserve high resolution and limit channel-level modeling. The key innovation in HsRu lies in the cross spatial reconstruction unit, which efficiently separates and reorganizes feature map information, balancing effectiveness and redundancy. Additionally, the skip connection components incorporate the CSRU to cross-combine spatial details between corresponding upsampling and downsampling layers, facilitating the fusion of information from various depths. This approach minimizes redundancy while maintaining a high level of information richness, ensuring optimal spectral reconstruction results.

To overcome the limitations of software demodulation in laser vibration measurement, a hardware demodulation scheme based on a field-programmable gate array (FPGA) is designed. The reference signal is synthesized using direct digital frequency synthesis and mixed with the Doppler signal. An orthogonal baseband signal is then generated through Butterworth low-pass filtering. Subsequently, a nonlinear error compensation algorithm is introduced to rectify errors. The CORDIC algorithm is applied for inverse tangent demodulation of the laser signal, while a phase detuning algorithm addresses signal hopping issues. Following data processing, the fast Fourier algorithm is used for time-frequency analysis of the vibration data signal. The proposed algorithm undergoes theoretical analysis and experimental verification, ensuring real-time vibration detection and meeting high precision requirements. This offers a more concise and efficient solution for laser vibration measurement systems.

Accurately identifying field weeds is crucial for selecting appropriate agricultural machinery and herbicides. This paper focuses on nine common weed species during the seedling stage in natural backgrounds of field vegetables, proposing a novel weed foreground segmentation method based on image processing and a new weed classification model using sparse representation classification (K_SRC). Initially, K-means clustering and threshold segmentation are applied to the a component of Lab color space to obtain binary weed images, followed by area filtering to remove interference. The binary weed image is then masked with the original image to extract clean weed images. Color features and mLBP texture features are extracted from segmented weed images to form a training sample set. KNN computes a local dictionary, and sparse coefficients are calculated using the OMP algorithm. Experimental results demonstrate a classification accuracy of 95.22% compared to four other algorithms. Additionally, this research provides new insights into the application of sparse representation in agriculture.

In the advancing fields of computer vision and deep learning, video dynamic object removal technology is an important research area. The goal of this technology is to remove specific dynamic objects from a continuous video sequence while filling in the removed background. This paper will explore a new video dynamic object removal method based on feature pyramid and point by point spatial attention mask extraction network. This method can detect dynamic objects and fill the holes caused by dynamic object removal in the video. We have designed and implemented an end-to-end dynamic target removal system based on the above methods. The system first applies RCNN for dynamic target detection and generates corresponding masks. Then, a generative adversarial network is used to complete video restoration of the removed target area. In addition, the system focuses on optimizing performance to ensure processing speed and result quality. The experimental results of this article indicate that this method has high accuracy and effectiveness in removing and repairing dynamic targets in videos. This indicates that this new visual technology method has broad rule-based capabilities in handling moving objects and repairing their backgrounds. In the future, this technology can be widely applied in various video processing and repair projects, such as video editing, security monitoring, and other fields.

The challenge in video super-resolution(VSR) is maximizing temporal correlation in video sequences and achieving precise inter-frame alignment. We proposed a bidirectional recurrent network integrating multi-scale deformable alignment with temporal attention fusion. We first designed a multi-scale cascading deformable alignment module, which captures spatiotemporal information at different scales through gradual refinement. It employs deformable convolutions for adaptive feature alignment to address complex motion and deformation issues. To effectively fuse the temporal information, we have added a temporal attention fusion module. This module dynamically weights the information of each frame using an attention mechanism. We carried out a comprehensive series of experiments on multiple standard VSR datasets. The outcomes indicate that our method achieved significant performance.

Aiming at the problems of insufficient utilisation of temporal information in the feature extraction process of existing multi-object tracking algorithms, as well as the problems of occluded trajectory localisation offset and trajectory loss during data association, we propose a 3D joint detection multi-object tracking algorithm based on instance-level temporal feature fusion and multi-stage association. The instance-level temporal feature fusion (ITFF) module utilizes attention mechanisms to merge instance-level features from previous and current frames, thereby alleviating the impact of object occlusion on the accuracy of trajectory regression. The trajectory correction (TC) module takes the trajectory regression box associated with the low confidence detection box as the output of the final trajectory to improve the precision of trajectory localisation. The trajectory recovery (TR) module further enhances the tracking performance of the model by fusing trajectory confidence to determine if a trajectory exists. The proposed algorithm is validated and tested on the KITTI dataset, achieving a high-order tracking accuracy (HOTA) of 76.01% on the test set, indicating better tracking performance and better robustness in occlusion scenarios.

Aiming to address the issues of low recognition accuracy, limited all-weather usability, and high missed detection rate in RGB-T multimodal pedestrian detection algorithms, this paper proposes an improved illumination-aware weight-based RGB-T pedestrian detection method. Firstly, the mid-term fusion strategy is adopted by changing the fusion point from after five layers of convolution to after the third layer of convolution. This ensures that the fused feature map contains both detailed and semantic features. Secondly, through learnable parameters, the illumination weights w_d and w_n are adjusted to more reasonably map illumination information to fusion weights, resulting in new illumination weights w_r and w_t. Finally, adjustments are made to the network structure of the semantic segmentation module by replacing two-stream semantic segmentation with fused single-stream semantic segmentation using new illumination weights w_r and w_t. The proposed method's detection ability is validated on a cleaned KAIST dataset. Experimental results demonstrate that compared to the original network model, our proposed method reduces average missed detection rate from 15.69% to 13.15% while improving mAP@0.5 by 4.1%, showcasing superior performance.

Through bionic learning of the human auditory system, sound source signal separation methods have made great progress. However, it is still difficult to use low-element arrays for signal separation in complex environments. In order to solve this challenge, this paper proposes a dual-element array sound source signal separation method based on compressed sensing theory based on the time-frequency masking technology, sound level difference and sound source azimuth correspondence map and greedy algorithm to achieve different azimuth sound source signals. of separation. Through theoretical analysis and simulation experiments, the effectiveness, accuracy and robustness of the proposed method are verified, and new ideas and methods are provided for further development in the field of subsequent signal separation, which has certain theoretical and practical application value.

Based on the Cloud-Edge system architecture proposed by the research team, this article investigates the method of Kalman filtering in acoustic emission signal denoising processing. Acoustic emission (AE) signals are small sounds generated when a material is subjected to stress or deformation, often disturbed by environmental noise, which reduces the clarity and usability of the signal. In order to effectively reduce the impact of noise on AE signals, this paper adopts the Kalman filtering algorithm. By dynamically modeling and estimating the signal, noise suppression and signal recovery are achieved. Through comparative experiments and analysis, it has been proven that the proposed method has significant advantages in noise reduction and signal fidelity of acoustic emission signals, and has good application prospects.

Equivalent field transformation (EFT) is a novel near-field testing technique. In response to the demand for measurement accuracy, numerical simulations were conducted to compare and analyze near-field acoustic holography and equivalent field transformation in a free field environment. The results showed that as the sound source frequency increased, the measurement accuracy of near-field acoustic holography decreased, while the equivalent field transformation always maintained a high measurement accuracy, Finally, an experimental environment will be established, and the theory of equal field transformation will be applied to near-field measurements of actual sound sources. The EFT results are basically consistent with the measured values, further verifying the correctness and effectiveness of this method.