This PDF file contains the front matter associated with SPIE Proceedings Volume 13063, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Developing the cash crop industry and helping rural revitalization is China's long-term development plan, applying image recognition technology to the field of cash crops, and efficiently realizing the picking and classification of crops. In this paper, three transfer learning methods are used to compare the transfer learning effects of AlexNet, VGG16, ResNet50 and other networks to select the most suitable model for fruit classification. ResNet50 migration training fully connected models can not only shorten the training time, but also achieve higher accuracy under the same iteration cycle. The accuracy of ResNet50 migration training fully connected layer model is 97%. The accuracy rate is improved by17%compared to VGG13.

To address the challenge of inadequate classification rate of the collected waste, the garbage classification image recognition model is designed based on machine vision technology, and after comparing the test effect of a variety of common network models after the collected 12 types of municipal household waste image dataset, it is found that Inception-ResNetV2 and Xception have significantly better recognition effects on the training and validation sets, with Inception-ResNetV2 performing better overall. This model was utilized as the pre-training model, and the model's output side was repeatedly tweaked for different parameters without affecting its internal stem and Inception-Resnet-A/B/C modules until the ideal model and related optimized parameters were identified. A confusion matrix was created to assess the accuracy and completeness of each form of garbage, and the model was shown to be better suited for identifying the majority of practical waste categorization applications such as biological, battery, cardboard, paper, and shoes.

Semantic segmentation is a fundamental problem in computer vision, aiming to assign predefined class labels to each pixel in an image. In order to reduce the annotation cost of semantic segmentation tasks, weakly supervised semantic segmentation has attracted increasing attention. Weakly Supervised Semantic Segmentation approaches commonly leverage class activation maps derived from classification networks. However, these classification networks primarily focus on regions that contribute the most to the classification task, resulting in imbalanced activation degrees and incomplete activation regions in the generated maps. To address these limitations, this paper proposes the Equilibrium and Integrity Correction Class Activation Map (EIC-CAM) algorithm. It introduces a Gaussian suppression strategy to enhance under-activated regions while suppressing excessively activated regions, thus overall enhancing the activation levels of target objects. Moreover, a category mining strategy based on pixel semantic correlation is employed to exploit the inter-pixel relationships and uncover more complete target regions. By implementing these two strategies, stronger activation maps with more complete regions are generated, leading to higher mIoU scores. Extensive experiments conducted on the Pascal VOC 2012 segmentation dataset demonstrate that, compared to several state-of-the-art methods, our proposed approach achieves a significant improvement in accuracy when utilizing only image-level annotations for training.

An accurate and fast fire smoke detection algorithm is urgently needed to solve the emergency linkage measures to prevent early fire spread and after the fire. In this paper, the depth matrix of motion difference is constructed for background modeling, the moving target is extracted using flame smoke motion characteristics, and the suspected area of flame smoke is obtained through corrosion and expansion. The unique characteristics of flame and smoke are then extracted, and the smoke flame recognition model is built using a BP neural network optimized by a genetic algorithm. The experimental results show that the algorithm can reliably eliminate the interference of moving vehicles and lights in the tunnel, detect flames and smoke, and generate an alarm, and that it can be applied to fire smoke detection in real-world scenes. Finally, the emergency linkage measures in the corresponding area are automatically initiated based on the position of the fire point via information interaction with the fire smoke detection and alarm system.

To address the problems of low accuracy and poor robustness of wildlife object Detection, this paper proposes an improved wildlife detection algorithm based on YOLOv7(You Only Look Once v7). The proposed algorithm introduces Deformable ConvNets v2 (DCNv2) and Wise-IoU (WIoU) to improve the model feature extraction and learning ability. In the self-built wildlife data set, when the Intersection over Union (IoU) was 0.5, the proposed algorithm was in36wildlife categories, the mean Average Precision (mAP) increased by 1.2 percentage points over the original YOLOv7percentage points, precision increased by 4.1 percentage points, and recall increased by 2.2 percentage points. Experimental results show that the proposed improved YOLOv7 algorithm performance is better, more can meet the actual wildlife protection work of animal detection and identification accuracy requirements, contribute to the wildlife local survey work, and save a lot of related resource cost, to a certain extent, promote the wildlife protection work.

To handle the existence of remote sensing image super-resolution reconstruction algorithms that are unable to fully utilize multi-scale information, insufficient feature extraction of remote sensing images and lack of learning ability of high-frequency information, This paper introduces a technique for enhancing the resolution of remote sensing images by employing multi-scale feature fusion and attention mechanism. The method firstly uses a single convolutional layer for preliminary feature extraction; secondly, a multiscale feature fusion-attention mechanism residual module (MFAM) is proposed in the nonlinear mapping stage, which uses three different sizes of convolutional kernels for feature extraction and fusion to make fuller use of the detailed parts of remote sensing images, and makes use of the serial-structured channels and the spatial attention mechanism to adaptively extract and enhance the high-frequency information; finally, sub-pixel convolution is used to realize up-sampling and complete the reconstruction for remote sensing images. we conduct comparison experiments of multiple methods on NWPU-RESISC45 and UC Merced datasets in this paper. The experiments in this article demonstrate that the proposed method has shown improvements across various evaluation metrics, resulting in superior super-resolution reconstruction results.

Substation is an essential part of the whole power system, which is burdened with the important functions of voltage and current conversion, electric energy receiving and distribution. At present, the operation and maintenance of substation in our country is mainly through regular inspection and maintenance of duty personnel, and there are many problems in related work efficiency and personnel safety. Therefore, with the development of robot technology, the power grid began to gradually replace some manual work with substation robots to solve the above problems. According to the basic requirements of the substation, develop the functional framework of the robot, develop the relevant hardware framework, understand the technology involved in the relevant hardware, and carry out the hardware selection. In this paper, the hardware framework is divided into three layers: main station layer, communication layer and implementation layer. Combining the advantages of traditional electric mobile robot and multi-degree of freedom robotic arm, an intelligent operating robot system with more comprehensive functions is designed. The positioning error of the moving robot is compensated step by step by image features, so that the operation unit at the end of the manipulator arm is in an ideal relative position to the operating object, and the operation unit is inserted into the operation hole by moving the manipulator along a fixed path to perform the corresponding operation. This method greatly reduces the calculation amount through step compensation and tentative movement. And the control accuracy also well meets the task requirements.

Multi-temporal remote sensing image change detection is an important research direction in the field of remote sensing, which has a wide range of applications in environmental monitoring, urban change, agricultural investigation, war symptoms prediction, military force deployment, military strike damage assessment and so on. In order to solve the difficult problem of change detection of some moving objects and small objects in remote sensing images, it is necessary to mine the global complex semantic information in the feature extraction stage, and the current research on this aspect is not in-depth. In this paper, a 3-branch TSPNet network based on Transformer structure and SKAttention mechanism is proposed. This model has advantages of high detection accuracy, full mining of global semantic information and high degree of refinement, which can realize fine detection and moving target detection. The model was tested on the CDD and Levi-CD datasets, and parameters such as F1 scores and mIoU outperformed other mainstream networks.

To find the suspect in forensic cases, a unique feature of shoe prints can be used. It is effective to distinguish people with shape of shoes and gait features demonstrated by many studies. To address the limitations stemming from relying solely on a single feature, such as complex equipment requirements and susceptibility to interference, this study explores a people identification system that leverages a BP neural network while integrating both shape and plantar pressure features. With 3 different approaches of data processing compared, an accuracy of 89% with no noises and 74%with10% gaussian noise. The result demonstrated that the system is effectiveness on people identification.

In the task of space target recognition, due to the limited conditions of image acquisition, the accumulated train dataset and test dataset have obvious differences in different poses, and the drastic changes of image representation caused by cross-view greatly restrict the performance of space target recognition. In this paper, cross-view space target recognition from multi-view stereo is proposed. The method is based on multi-view stereo to estimate the image depth information, which is fused with texture information of 2D image, and finally uses the 3D CNN to classify the image. The network learns and inferences depth information without the supervision of depth map truth value, which adds 3D information such as geometry of space objects to the classification task, and improves the performance of target recognition. The experimental results on the self-made space target data set prove that the proposed network MVS-STRNet is more efficient on the cross-view data set.

Visibility prediction affects travel safety and is a major concern for highway and airline authorities. However, current visibility prediction methods based on video data usually capture some features in static images, which are relatively simple and are not commonly used for continuous variation information of video data. This study proposes a visibility prediction method based on the spatio-temporal variation features of a video. Through the spatial division of the images from the video, the optimal spatial distribution of different features in the image is mapped using the correlation matrix, and the spatio-temporal variation information of the video is extracted based on the short-term stationary features of fog spatio-temporal evolution. Finally, the relationship between the spatiotemporal variation features of the video and visibility is constructed by combining the time-frequency localization features of the wavelet transform and the self learning ability of the neural network. Compared with the predicted results of traditional static image features, theR2, RMSE, and MAE of the visibility prediction results were improved by 0.1698 and reduced by 58.4142 and 20.0427, respectively. In addition, the R2, RMSE, and MAE of the prediction results based on wavelet neural networks can reach 0.9817, 28.3365, and 19.2098, respectively, compared to current mainstream prediction methods. This fully proves the scientificity and effectiveness of the method, which can be applied to the accurate monitoring of visibility in fog weather based on video data.

This paper compares the effects of SE-Net and the improved CBAM attention mechanism, and proposes a CTSA (Channel and Temporal and Spatial Attention) attention mechanism that adds temporal attention for temporal features, and further proposes a TSA (Time-domain and Spatial Attention) attention mechanism that focuses on time-domain on the basis of comparing the effects of multiple attention mechanisms. The experimental results show that in the field of behaviour recognition, the TSA attention mechanism designed in this paper can achieve better recognition results when using a lighter weight network structure.

Video quality assessment (VQA) plays a crucial role in video transmission and storage to ensure optimal video compression effect. In existing research, deep neural networks are commonly employed to align with the complexities of the Human Visual System (HVS) and achieve superior performance. However, due to the complexity of the HVS, many characteristics of HVS have not been considered, resulting in a suboptimal fit to the HVS. To address this issue, this paper proposes a novel VQA method grounded in the HVS. Notably, our method introduces brightness information from video frames as embeddings, thereby enhancing the model's alignment with the HVS. Additionally, we identify that feeding every single frame into the network for feature extraction leads to a lot of redundant information. To improve efficiency while maintaining comparable performance, our method samples the video frames at an interval of eight, resulting in an eightfold improvement in efficiency. The effectiveness of our method is validated through extensive experiments conducted on four mainstream VQA datasets. Ablation studies further corroborate the effectiveness and efficiency of our method by demonstrating the positive impact of incorporating brightness information into VQA.

Glands and polyp automatic segmentation is the basis of artificial intelligence auxiliary diagnosis of colorectal cancer. Considering that polyps have different sizes, shapes, colors, textures and blurred boundaries with surrounding mucous membranes, an enhanced receptive field and boundary network (ERBNet) was proposed for polyp segmentation. Firstly, res2net50 is used as the backbone network to extract image features. Secondly, a novel receptive field enhancement module (REM) is designed to obtain both low-level features with boundary details and high-level features with global semantics. Finally, the boundary enhancement module (BEM) is used to enhance the edge contour information and improve the segmentation performance of the network for small targets and polyps with fuzzy boundary organization. ERBNet was tested on five polyp datasets and achieved good results in mean Intersection over Union (mIoU) and mean dice coefficient (mdice).

Electromagnetic waves undergo attenuation as they propagate through space, and this attenuation is influenced by the properties of the objects they penetrate. This characteristic presents opportunities for the application of electromagnetic waves in computed tomography (CT) technology. Through scanning an object from various angles, it can be divided into multiple subregions. By combining the received electromagnetic field strength values obtained during the scanning process, a linear equation system can be formulated, which establishes a relationship with the attenuation coefficients of each subregion. Solving this equation system enables the identification of internal structures and the successful generation of CT images based on the values of the attenuation coefficients.

Face is an important source of information for interpersonal communication and recognition, so facial reconstruction technology has always been a research focus in the field of computer vision. We present a deep learning-based algorithm for accurate three-dimensional face reconstruction from a single-view video stream. The proposed method processes the input video stream frame by frame, extracts facial region information, and constructs a facial reconstruction network that utilizes 3D Morphable Models to reconstruct the precise geometric shape of the face. Additionally, we design a multitiered loss function, including low-level pixel consistency loss, facial landmark loss, and high-level identity loss. Furthermore, these multi-tiered losses are utilized as weak supervision signals to guide the supervised learning of the reconstructed face, thereby enhancing the quality and accuracy of the reconstruction.

Video anomaly detection is extensively utilized across a variety of domains including public transportation, industrial production, city management, and military fields to mitigate risks and bolster enhance safety. To tackle the challenges associated with video anomaly detection in intricate environments, we propose a light but efficient framework that builds upon future frame prediction techniques. Our framework incorporates Convolutional Long Short-Term Memory (ConvLSTM), masked convolution, and attention mechanisms to enhance the detection accuracy. Furthermore, to simplify the model's complexity, we replace the convolutional layers in the network with depthwise separable convolutions (DSC). Through evaluation on public datasets such as CUHK Avenue, UCSD Peds1, and UCSD Peds2, our proposed network model exhibits both high accuracy and real-time performance.

Colorizing grayscale images automatically has consistently posed a formidable challenge, and in recent years approaches using deep neural networks have become mainstream techniques. However, the results of colorizing these images remain unsatisfactory, ignoring color richness and structural consistency. As a result, we propose an image coloring method based on the combination of a sparse attention mechanism network and a color distribution predictor. Firstly, the color distribution predictor uses anchors to predict the color distribution of different regions or pixels in the image, so that the model can better understand the color relationship of different regions in the image, and make the coloring results more natural and consistent with the real-world color distribution. The sparse attention-based Transformer network is then used to generate a low-resolution coarse coloring by reference to the sampled anchor color first, before upsampling it to a high-resolution image. Sparse attention not only accelerates the training and inference process of the Transformer model, but also improves the coloring quality as well as preserves image details. The results show that our method achieves significant superiority, reducing computational complexity, improving efficiency, and producing more realistic color images with better coloring results.

Plant leaf disease few-shot classification can learn from a few samples to recognize the disease class automatically, which can protect the agriculture yield and quality. However, the support set of previous works still very big, which introduce 50 samples per class for training, which is still relatively large. Therefore, the training way of previous work is not in line with the few-shot classification. In this work, we propose a new network SProtoNet, which introduce the self-supervised learning and our proposed Prototype feature extractor to learn a more discrete feature space. And in order to test our proposed SProtoNet, we introduce the typical 5way1shot and 5way5shot to evaluate. Furthermore, we also propose a new baseline, and our network can achieve the state-of-the-art performance in the Plant village dataset.

In this work, we delve into the implementation approaches of two different classic attention mechanisms and the effects of two distinct methods for adding attention mechanisms in U-Net image segmentation tasks. Based on experiments, we further conclude that in the context of image segmentation tasks in computer vision, attention mechanisms, while achieving weighted channel information, can have a negative impact on the backbone network of conventional convolutional feature extraction. However, they can effectively compensate for the shortcomings when it comes to encoding high-dimensional semantic information in the encoder-decoder model structure.

Semantic segmentation of remote sensing urban scene imagery is a dense prediction task, which has been applied to the land-cover or land-use category. However, the dimension of remote sensing image is huge, which will result in the huge computation cost. In order to reduce the computation cost, a common method is to design a lightweight decoder to achieve a good trade-off between accuracy and computation cost. For this purpose, we design a lightweight transformer-based decoder GDformer. The GDformer consists of our proposed Global Value Transformer and Dynamic feature fusion module. The Global Value Transformer can extract the global semantic feature and the Dynamic feature fusion module can fuse the local feature and global semantic feature dynamic to capture the local-global context with a good trade-off, and the local-global context has been proved is necessary for the semantic segmentation of remote sensing. Extensive experiments prove that our proposed method can achieve a good trade-off between and state-of-the-art performance.

According to the data from the China National Bureau of Statistics, traffic accidents caused by fatigued driving account for more than 40% of large-scale traffic accidents [1]. Therefore, how to effectively detect and accurately warn fatigue driving is a very important topic in the field of transport safety. In this paper, we propose a fatigue driving warning device based on a "facial recognition fatigue driving detection system", which improves the accuracy and effectiveness of fatigue driving warning through five steps: facial information pre-entry, face detection, state recognition, fatigue assessment, and warning. Equipped with a Raspberry Pie 4B camera, the device digitizes the collected information using the OpenCV and Dilb libraries and determines the driver's fatigue level by calculating the Eyelid Aspect Ratio (EAR), the Mouth Opening Ratio (MAR), and the Euler Angle of the head. In addition, the warning device also grades different levels of driver fatigue and provides corresponding warnings, reminders, and treatment measures, which are of great significance in reducing traffic accidents and ensuring traffic safety.

Surface quality inspection is a crucial step in industrial manufacturing. However, it is challenging to collect an adequate quantity of abnormal samples in practice. Supervised methods require sample annotation, which is costly, so unsupervised methods that are high-speed and low-cost are more suitable for industrial applications. Among the current unsupervised methods, embedding-similarity methods haven shown excellent performance, but most of them do not preprocess the images and directly use convolutional neural networks to extract image features. While in actual scenarios, image can have some degree of offset or rotation due to machine variations. Therefore, this paper proposes a new patch distribution framework for anomaly detection, specifically a novel image alignment module is proposed to enhance the utility of the model. Image alignment reduces the dense distance between pixels during training, enabling more precise learning of the feature distribution of normal samples and reduce false positives during testing. In addition, in the feature extraction stage, the middle layer of the network is selected to extract features and establish embedding connections. This not only enhances the model’s precision but also reduces its memory requirements. Experiments on the publicly available datasets MVTec AD and BeanTech AD show that our proposed new framework achieves better performance than other baseline models.

To ensure the safe, stable, and reliable operation of substations and improve the intelligent anomaly monitoring ability, this paper proposes an algorithm for detecting abnormal behaviors and equipment defects of substations based on an attention mechanism and feature fusion network. This algorithm designs the feature fusion structure of BiMFPN. It constructs multiple green channels of shallow input to deep output feature fusion, thus fusing a more extensive range of practical features without increasing the cost. At the same time, lightweight CBAM is added to the fusion network to effectively combine the extracted spatial location features with the rich semantic information in the image. In addition, the algorithm also adds a new detection head, which effectively improves the detection performance of multi-scale targets and is more sensitive to small targets. After experimental tests, the mAP and F1 scores of the algorithm are 69.3%and 71.7%, respectively, which are 9% and 8% higher than that of YOLOv5m. The experimental results show that this method can effectively detect abnormal behavior and faulty equipment around the substation and improve the intelligent detection level of the substation.

The Blockformer speech recognition model has recently been proposed as a state-of-the-art (SOTA) model ontheAishell-1 Chinese speech dataset. This model exhibited significant improvements in character error rate (CER) when compared to its baseline, Conformer. The key improvement of Blockformer is the addition of the Squeeze-and-Excitation (SE) block on top of Conformer, which enables better utilization of the information contained in each Conformer block. In our study of Blockformer, we identified scope for improving its block information extraction method. To this end, we used the attention mechanism to enhance the SE block's efficacy in squeezing block information. And we enhanced the model's structure in attention inference mode to align more effectively with the training structure. Under the four inference modes, namely attention, attention rescoring, ctc greedy search, and ctc prefix beam search, the CER reaches 4.67%, 4.43%, 4.75% and 4.75%. All of these rates are at the level of Blockformer or exceed it.

The current vision-guided relevant target object pose recognition methods applied in intelligent robots have the problems of poor accuracy and simple application scenarios, so the study optimized its network on the basis of the de-entangled pose network of coordinates and experimentally verified it. The experimental results show that in the analysis of the network test results, the average value of the improved pose recognition estimation network is 92.05%, which is significantly higher than the 89.86% of the original de-entangled pose network of coordinates, and is significantly higher than the other comparison methods. In the analysis of the network output visualization results, the study uses the data with 5px as the threshold to visualize and display, at which time the accuracy of all objects exceeds 90%. And its overall success rate reaches 92% when it is practically applied in intelligent robots. Comprehensively, the improved target object attitude estimation network proposed by the study has the effectiveness and practicability, which can effectively improve the degree of grasping intelligence of intelligent robots.

Conformer has demonstrated outstanding performance in the field of speech recognition, representing a highly effective improvement over the Transformer model. Subsequently, research on the integration of encoder blocks emerged, such as Blockformer. In this paper, we investigated the integration method of Conformer's encoder blocks and proposed a more efficient approach for leveraging information from multiple blocks, which we refer to as Convolution-Enhanced Channel Attention Block Integration. We achieved a Character Error Rate (CER) of 4.58% on the Aishell-1 dataset.

Under the rapid development of today's society, our life is more and more inseparable from intelligence. At present, through the widespread popularization of intelligent technology and the gradual deepening of research projects, image analysis technology has been applied in many fields, and for image analysis technology to be applied to the field of interaction design can further enhance the sense of intuitive and intelligent user experience, and can also provide a strong and powerful support for interaction design. This paper provides a comprehensive analysis of the research field of image analysis and the application of interaction design, and identifies the advantages of the combination of the two, as well as makes further optimization for the shortcomings.

The safety of concrete structures is crucial for the sustainable development of infrastructure. However, accurate detection of concrete surface cracks has always been a challenging task. With the continuous development of deep learning technology, image-based concrete crack detection has made significant progress in the field of automation. However, running efficient deep learning models on embedded devices remains challenging due to their limited computational resources and storage capacity. To address this challenge, a surface crack image classification method based on multi-model voting is proposed in this paper to enhance real-time performance on embedded devices. In this study, three classical deep learning models were selected: InceptionResNetV2, VGG16, and Xception. By combining these models, we leverage their unique advantages in image feature extraction, constructing a multi-model fusion classification system. During the classification process, a voting mechanism was employed, ensuring the reliability of the final classification results. Experimental results demonstrate that compared to single-model approaches, our multi-model method achieves significant improvements in both classification accuracy. The contribution of this research is not only reflected in technical innovation but also provides a feasible solution to the practical problem of automated concrete surface crack detection.

In recent years, the existing target detection algorithms have difficulties in detecting small targets, such as low target resolution, large interference in complex scenes, and lack of large and complete small target detection data sets. To solve the above problems, this paper proposes a new Yolov5 based target detection improvement model E-b-yolov5. This method first preprocesses the data, and then adds ECA to the last layer of Backbone structure by changing the basic network of Yolov5, This lightweight module effectively avoids dimension reduction. Secondly, CIoU is used as the loss function of frame regression to achieve high-precision positioning. In addition, Bifpn (Efficient Det) feature fusion is added to enhance the feature representation ability for small target detection, and the removal of single input side nodes reduces the amount of calculation. Finally, the model is distilled to increase the recall and accuracy of the model. Compared with the original yolov5 method, E-b-yolov5's mAP on the validation subset of the aerial photography dataset VisDrone 2019 DET reached 41%, 8% higher than YOLOv5's benchmark network; On the test subset, mAPreached33.4%, 4.3% higher than Faster R-CNN. research on small object detection to improve the detection effect of small object technology and extend its application scope in practical scenarios.

In order to improve the safety of nighttime car driving, this paper proposes a lightweight image fusion network based on depth-separable self-encoder combined with attention mechanism for the problems of insufficient extraction of highdimensional spatial information, loss of gradient information and edge blurring in infrared and visible image fusion, and introduces Inverted Residual Block into the encoder, using depth-separable Inverted Residual Block is introduced into the encoder to reduce the number of parameters used in feature extraction, shorten the convolution time, invert the residual structure and linear bottleneck to effectively preserve the high-dimensional spatial features, and rescale the extracted features in the attention mechanism module to enhance the image information, and finally use L1 regularization in the fusion layer to fuse the image features and reconstruct the image output through the convolution layer of the encoder. The gradient loss function is introduced in the training network to preserve the gradient information of the image. The experimental results show that the method in this paper is more consistent with the human visual system in subjective evaluation and has significant improvement in both objective evaluations with 22.89%, 56.65%, 5.35%, 15.39%, 32.54% and 3.81% improvement in spatial frequency, standard deviation, gradient-based fusion performance, difference correlation sum, visual information fidelity and multi-scale structural similarity, respectively.

At present, with the continuous development of computer vision technology, people's requirements for images are no longer limited to the degree of visual clarity, but pay more attention to the comprehensiveness of image information. Compared with the limited field of view of the traditional perspective camera, the fisheye camera has a blind-free field of view and can obtain rich visual information. However, the wide field of view of the fisheye camera causes large-scale variation and distortion problems, which creates a great challenge to the feature extraction of traditional object detection algorithms. According to the characteristics of fisheye images, we adopt the current state-of-the-art lightweight YOLOv8s model as the base model and introduce Sphere Convolution (SphereConv) to deal with the complex distortion characteristics of fisheye images. The proposed Sphere-YOLOv8s model effectively improves the distortion problem of fisheye images. The experimental results show that the Sphere-YOLOv8s model on the VOC-360 dataset obtains an excellent improvement in object detection compared to the original YOLOv8s model. Among them, Precision improves 12.5%, Recall improves 14.5%, mAP@0.5 improves 16.7%. mAP@0.5:0.95 improves 20.5%.

In the field of computer vision, video semantic segmentation algorithms are fundamental tasks and have been widely studied. However, existing algorithms have room for improvement in terms of segmentation accuracy and segmentation time. To address this problem, this paper proposes a memory network segmentation algorithm based on a multi-scale spatio-temporal polarized attention mechanism. The algorithm consists of three parts: a multi-scale memory network, a large kernel decomposition convolution module, and a spatio-temporal polarized attention module. The multi-scale memory network is used to store historical frame feature information at different scale levels. The features are transformed by the large kernel decomposition convolution module, aiming to achieve a balance between expanding the sensory field and reducing computational effort. Finally, the spatio-temporal polarized attention module captures interframe similar feature information and performs feature fusion on the inter-frame information to improve segmentation accuracy. The algorithm achieves a segmentation accuracy of 75.01% and 72.98% on the Cityscapes dataset and Camvid dataset, respectively. The experimental data proves that the algorithm achieves a more advanced level in terms of segmentation accuracy and speed.

In order to improve the chart recognition speed and the recognition success rate of the large field-of-view high-sensitivity star sensitizer, an improved artificial bee colony algorithm combined with the three anti-perturbation parameters is proposed as a chart recognition method. The method is based on the improved artificial bee colony algorithm to find the optimal paths for all the stars in the main star feature region; then, the optimal path length is used to find addresses in the navigation feature library, and the interstellar angular distances of the first three navigation stars in the optimal paths, d12/d23 , are taken as the main recognition features for matching and identification; finally, the three perturbation parameters are taken as the auxiliary recognition features to reduce the redundancy of the matching and to serve as a validation of the recognition.

Real environments exhibit higher frequencies of complex scene conditions, characterized by their inherent variability. The loss of detail, loss of contrast, blurred picture and other degradation phenomena in the captured images, which may affect visual perception and the performance of computer vision tasks. This paper studies the image enhancement problem under complex scenes to improve the performance of computer vision tasks. This paper proposes a continuous learning-based deep neural network for enhancing images over different categorical scenes. The proposed model incorporated a Cross-Attention Module (CAM) into the image enhancement model and introduced an image compression reconstruction model within the continuous learning framework. The experimental results demonstrate that the proposed model demonstrates its accurate performance and its advantages than the baseline methods.

Transformer is a new type of deep learning model that relies on attention mechanism. This study investigated a way of correctly identifying both the accent and gender of audio with high accuracy and preciseness. Thorough trainings were conducted with 1000+ hours of audio recordings, our model classified the gender and the accent. Multitask refers to the capability of a program to perform multiple tasks or processes concurrently. The model we finished not only correctly classify the accent of the audio but also gender.

Biomedical named entity recognition is of great significance in the field of natural language processing.Deep learning approaches are mainly used at present, and the BERT-BiLSTM-CRF model is one of them. Although the BiLSTM structure in this model can capture bi-directional long-distance dependencies in sentences, there is the problem that the model performance can easily reach the bottleneck by simply adjusting the hyperparameters or changing the network structure, and the quality of its recognition results can hardly be further improved. To address the above problems, this paper adds the label correction process based on deep reinforcement learning to the BERT-BiLSTM-CRF model to form the BERT-BiLSTM-CRF-DRL model. The intelligent body Agent based on deep reinforcement learning is used as a label corrector, and the label correction threshold is set, so that the uncertainty in the annotation results of the BERT-BiLSTMCRF model is greater than the label correction threshold. The labels with uncertainty greater than this threshold in the results are processed for correction. The experimental results show that the F1 values on the datasets BC4CHEMD, BC5CR-CHEM, and NCBI-DISEASE reach 91.07%, 94.02%, and 89.56%,, which are all improved compared with the existing advanced biomedical named entity recognition models.

There are two major problems with traditional recommendation models: firstly, when user preferences change, the recommendation model does not change accordingly; Secondly, the limited ability to handle abnormal data leads to a lack of diversity in recommendation results due to the singularity of recommendation results. User behavior data has been a hot topic in the field of recommendation research in recent years, and it can be used to mine deeper user information. To extract user preferences from comment text and user source information, a UICTM model is proposed. Based on this, LDA technology is used to obtain the UIFT algorithm, which integrates time factors to optimize the UIFT algorithm. CF, TMF UIFT+was used for experimental data analysis and comparison. The results showed that the UIFT+algorithm outperformed other recommendation algorithms overall in terms of mean square error (MSE) and recommendation degree (ACC).

In augmented reality, the visual consistency of virtual-real fusion is a difficult problem. For the problem of detecting and generating information about the shadows of virtual objects in complex scenes from a single viewpoint, a method is proposed to achieve the consistency of lighting on virtual objects based on a panoramic view, which makes up for the shortcomings of shadow information collection in complex real environments from a single viewpoint. First, the T-ORB algorithm is used for image stitching to generate the panorama. Then, the shaded information of the real environment is fully collected based on the panoramic view of the complex environment. Finally, the registration of virtual objects is completed according to the transformation matrix and rendered in combination with the shadow body. Experiments prove that the method proposed in this paper achieves shadow rendering of virtual objects, which greatly improves visual consistency and has obvious advantages in rendering time.

Brain tumor is one of the common diseases of nervous system. Because of the different size and location of each tumor region in the brain tumor image, the traditional brain tumor image segmentation method has poor segmentation effect at the edge of each tumor region. In recent years, deep learning has been applied to medical image segmentation this paper proposes an improved attentional Unet model. Firstly, the complex 3D brain tumour image segmentation is decomposed into simpler 2D segmentation tasks. Secondly, the network model is tested and compared with the segmentation results of the classical segmentation network, the network in this paper has a good segmentation effect.

The popularity of the internet and the development of mobile internet has led to a rapid growth in internet users and smart devices and a rapid increase in internet traffic consumption[1]. Through an in-depth study of user video clicks and VBR encoded video[2] bit rates, we propose a new and effective push scheme. Based on the storage and channel configuration conditions in the hybrid network, as well as the VBR-encoded video characteristics, a push formula is given for video split, single programme and multiple programmes; with the goal of saving on-demand traffic and ensuring user storage prerequisites, the problem is dynamically planned for the rotation and push scheme selection problem for video split, single programme and multiple programmes. In this paper, the efficient content distribution scheme is based on the premise of guaranteeing memory storage, the goal of saving traffic, and the comparative reference quantities of push, rotation distribution and storage benefits. The dynamic planning problem is analyzed and optimized to a 0-1 backpack problem to find the push and rotation slice numbers and quantities. By improving the efficiency of Internet video content distribution and reducing distribution traffic, the efficient content distribution solution can effectively alleviate network congestion and the rapid growth of CDN investment, thus improving user experience.

The cognitive decline caused by Alzheimer's disease (AD) is closely related to changes in the hippocampal structure captured by magnetic resonance imaging (MRI). To accurately analyze the morphological changes in the hippocampus caused by AD, it is necessary to extract the landmarks on the hippocampal surfaces which can ensure accurate registration between the different hippocampal surfaces, to eliminate morphological measurement errors caused by factors such as image acquisition, age, and gender. In this paper, we propose a method that can automatically extract the intrinsic landmarks of the hippocampal surfaces. More specifically, we use the eigen-graph of the hippocampal surfaces to extract the intrinsic and unified landmarks of all the hippocampal surfaces and calibrate along the longitudinal direction of the hippocampus surfaces along with the landmark correction for a few irregular hippocampal surfaces. Finally, we apply the extracted landmarks to the landmark-based registration algorithm, and the results show that landmark-based registration using our extracted landmarks can improve the accuracy of AD classification.

Visually secure image encryption has become a hot topic in the information security field. However, the embedding capacity of the existing visually secure image encryption schemes is too limited. To solve this problem, a visually secure image encryption scheme using parallel compressive sensing (PCS) is proposed in this paper. Firstly, the secret image is preprocessed by using global random scrambling. Secondly, the scrambled secret image is compressed by PCS. Thirdly, the compressed bitstream is encrypted through a XOR encryption operation. Finally, the encrypted bit stream is embedded into the low bit-plane of the carrier image. The experiment results demonstrate that the proposed scheme significantly increases the embedding capacity in comparison with the current visually secure image encryption schemes.

Quantization parameter (QP) is an important factor to determine the bitrate in video coding, especially in the case of limited bandwidth, how to choose the quantization parameter reasonably to achieve the actual bitrate as close as possible to the target bitrate is particularly important. In this paper, Differential Evolution (DE) algorithm is used to assign quantization parameter for Coding Tree Units(CTUs) in each frame of 360-degree video based on High Efficient Video Coding (HEVC). The experimental results show that compared with HEVC reference software HM-16.20, the proposed algorithm can provide 1.86% bitrate savings and 0.07dB improvement in Peak Signal to Noise Ratio (PSNR).

Small Form Pluggable (SFP) optical modules have become the mainstream optical module packaging because of their advantages such as small size, low cost, and high reliability. Sensitivity testing is an important step to verify the performance of optical modules. The experimental platform utilized to test the sensitivity of SFP optical modules currently has some issues, including complex equipment structure, high volume, and poor testing efficiency. This paper proposes and implements an evaluation board that can be used to enhance the experimental platform for the SFP sensitivity test to address these issues. The evaluation board's merits include its compact size, strong expansibility, and ease of use. The experimental results demonstrate that using the evaluation board to construct an experimental platform for SFP optical module sensitivity test significantly reduces the average test time and triples the overall test efficiency.

Due to long term development by waterflooding in the unconsolidated sandstone oilfields, high permeability flowing channels or “large pore paths” are likely to form between some injection and production wells, resulting in the channeling of injected water as well as the worsening of waterflooding effectiveness. Therefore, how to quickly and accurately identify and predict the direction of high permeability flowing channels is very significant for tapping the remaining oil and designing the scheme of profile control and water plugging. According to the deficiencies in some testing methods which are commonly used in field to discriminate high permeability flowing channels, a convenient, efficient and economic approach to predict the distribution direction of high permeability flowing channels is put forward in the paper, which base on system analysis theory can fully dig out implicit information behind the dynamic data. Moreover, the validity of this method has been verified by using theoretical model and actual production data.

Object detection is a challenging task in computer vision that involves predicting both the class and the location of objects in an image. Most existing methods rely on convolutional neural networks and hand-crafted modules, such as anchor boxes and non-maximum suppression. Recently, a novel end-to-end approach called DETR was proposed, which uses a transformer encoder-decoder structure to model object detection as a set prediction problem. However, DETR suffers from some limitations, such as poor performance on small objects and slow convergence speed. In this paper, we propose FF-DETR, a feature-fusion detection transformer that improves the performance and convergence speed of DETR-like models. FF-DETR introduces three feature fusion modules: (1) Contour Fusion FPN, which fuses multi-scale features using self-attention and deformable convolution; (2) Position-Content Query Fusion, which initializes the content query features by fusing the position query features and the encoder output features; and (3) Global Decoder Layer Fusion, which fuses the outputs of each decoder layer and updates the position query features iteratively. We conduct experiments on the COCO dataset and show that FF-DETR outperforms DETR and other variants in terms of accuracy and efficiency.

Target tracking in computer vision is a challenging task that requires both accuracy and robustness, especially in the presence of occlusions and complex backgrounds. To tackle these challenges, this paper proposes a novel tracking approach based on correlation filters, which incorporates reliability assessment and re-detection modules. The proposed approach leverages enhanced color histogram scores in combination with correlation filter response maps to improve the accuracy of target localization. By considering metrics such as maximum response value, average peak-to-correlation energy, and peak-to-sidelobe ratio, the reliability of the tracking process is evaluated, providing insights into its performance and stability. In cases of tracking failure, a re-detection module is activated to reacquire the lost target and resume the tracking task. This module dynamically adjusts and updates the model to enhance the accuracy and robustness of the tracking process. Extensive experiments are conducted on popular benchmark datasets includingOTB50, OTB100, as well as custom datasets. The results demonstrate the superiority of the proposed approach, particularly in challenging scenarios with occlusions and complex scenes, outperforming baseline trackers and other state-of-the-art tracking methods.

In the realm of art education, the assessment of color blending and color differences has traditionally been a manual process, often subject to human subjectivity and variations in visual acuity. This paper delves into the application of computer vision techniques, a significant branch of artificial intelligence, to address these challenges. By leveraging the high precision, objectivity, and fast processing speed of computer vision, this study aims to revolutionize color inspection in art education. The research encompasses various stages, including art image acquisition, preprocessing, color difference calculation, and feedback provision. A focal point of the study is the extraction of features from art images and the exploration of classification algorithms for online color difference assessment. Through the application of a data modeling method based on the Morlet wavelet kernel SVM, the research achieves an accuracy rate of 88.67% in the classification of fine art images. While promising, the study also identifies areas for further research, including the quantification of color discrepancies in intricate art images and the enhancement of the classification algorithm's operational speed.

Railway catenary systems are essential equipment for electrified railways, and their safe and stable operation directly affects the regular movement of trains and the safety of passengers' lives and property. Foreign objects are one of the main threats to the safety of the catenary system, making the detection and classification of these foreign objects a challenging task. This paper proposes a method that cascades the YOLOv5-lite and DenseNet-121 networks, carrying out detection and classification in separate steps. This approach facilitates detection in high-resolution scenes with complex backgrounds, thereby reducing computational overhead and increasing accuracy. Such a method ensures that the detection network does not overly focus on the minute details of the foreign objects in the catenary system.

Privacy preserving machine learning (PPML), which focuses on training machine learning models while preserving privacy, has attracted many research interests in the recent past. This paper investigates recent research on privacy preserving machine learning (PPML) techniques and their energy applications. PPML allows training models while protecting sensitive data. We categorize approaches into federated learning, homomorphic encryption, differential privacy, and secure multi-party computation. After explaining each method’s principles, we summarize research applying them to energy tasks like forecasting and anomaly detection. Comparative analysis identifies trade-offs between privacy guarantees, costs, and accuracy. Encryption-based techniques provide strong privacy but impact efficiency. Federated learning is more efficient but has weaker protections. Ongoing challenges include the performance of cryptography, calibrating noise, and developing robust aggregation rules. The survey offers a comprehensive reference on the state-of-the-art in PPML for energy, guiding further advancement of practical systems.

In this study, we aim to evaluate the applicability of the VGG 19 architecture-based neural network for autonomousdriving tasks. The objective is to demonstrate the effectiveness of employing this neural network to solve the problemand to compare the identification quality. Experimental methodology is utilized to compare the efficiency of different neural networks. During the experimentation, we observed a disparity in the efficiency of VGG with and without theinclusion of the attention mechanism. Presently, a variety of machine learning methods are increasingly used to addressdiverse problems. Image recognition holds particular significance among these challenges. The practical problemunderinvestigation in this study pertains to the identification of roads with varying curvatures and their correspondence tothevehicle’s steering angles for the purpose of autonomous vehicle control. This paper explores multiple types of neural networks commonly used for image recognition. Previous research has demonstrated the high efficiency of these neural networks in solving various tasks, while this evaluation focuses on the applicability of neural networks in solvingtheautonomous driving problems and a performance comparison of these methods. The experimental results indicatethat VGG 19 achieves the highest accuracy of all neural networks.

In this paper, position and orientation measuring method of auto-landing based on pure visual information deep learning method combined with traditional methods was proposed. The technical methods and principles used in the calibration of long-distance imaging camera, detection and tracking of runway target, point and line feature extraction of runway target and high-precision pose solution are described in detail, and relevant experiments were carried out in the paper. The results shown that the proposed method could realize the pose measurement with lateral deviation error better than ±2m, vertical height error better than ±1m, azimuth and pitch Angle error better than ±0.4° during the aircraft landing.

In response to the call for sustainable development and seeking new break-throughs in achieving green transformation across the entire industrial chain and management processes, this study introduces the XGBoost regression method. It constructs a machine learning model using 140,566 historical data points collected from the data acquisition system. Additionally, it incorporates local climate factors such as average temperature, weather conditions, maximum temperature, minimum temperature, and wind speed. The model is thoroughly explored and analyzed. Experimental results confirm that the model developed in this study exhibits commendable accuracy and stability. It provides a strong theoretical foundation for achieving cost reduction and efficiency enhancement in factories, aligning with the principles of green development.

In some IoT applications, the sensing layer often collects images and various temporal data. Usually, some temporal data represents sensitive or private information. In order to protect the temporal data, and at the same to reduce data traffic transmitted through the Internet of Things, this paper proposes a secure transmission method for sensitive data based on Arnold transform and steganography technique. In this method, an Arnold transform which is usually used in scrambling images is applied to encrypt sensitive sequential data. It is achieved by re-organizing the position of data elements in a matrix, not necessarily a square matrix. Then, the sensitive data in the matrix is embedded row by row in a carrier image using a novel data hiding technique, namely, capacity adaptive algorithm. Through the above dual processing, the sensitive data is less likely to be detected by a third party during transmission. The experimental results indicate that the method achieves high performance in safeguarding the security and privacy of sensitive data. In addition, the method offers numerous benefits, including high efficiency and scalability in data hiding capacity, and low computational cost, making it widely applicable for practical data transmission.

The challenges of autonomous driving stem from its exposure to complex environments and stringent safety requirements. End-to-end autonomous driving entrusts the entire driving task to a single large model, allowing the model to directly derive control signals from sensor inputs. Although this approach has shown promising results, many end-to-end models require significant human effort for training through imitation learning. Therefore, we aim to establish a reliable and cost-effective model by combining existing image recognition models through transfer learning. The selected model for this study is Inception V3. Based on this, experiments were conducted by adjusting the learning rate, freezing certain layers, and modifying the fully connected layer, resulting in improved model performance. This enabled better planning of driving paths.

This research offers a denoising model that organically blends Generative Adversarial Network (GAN) and Kernel Prediction Network (KPN) to address the issue of probable loss of fine details in denoised Monte Carlo pictures. To preserve image details during the denoising process and overcome the denoising network's inadequacy in image feature extraction, the model introduces a Multi-Scale Feature Module to expand the receptive field of feature extraction, allowing the model to extract image features at different scales. And employs Haar Wavelet Transform to suppress high-frequency noise in images, hence increasing the model's denoising performance. The suggested network model efficiently lowers noise in diverse image scenes, improving quality metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).

Vein recognition has received increasing attention in recent years. Currently, the deep learning such transformer show robust capacity of feature representation and successfully applied for vein recognition task. Despite of its advances in vein recognition, but it may fail to capture the 2D structure and spatial local information within each patch for image-based vision tasks. In this paper, we propose a novel Convolutional neural network based transformer for vein recognition, termed VeinCnnformer to take advantage of convolutional operations and self-attention mechanisms for improving representation learning. Our VeinCnnformer roots in the Mixed Convolutional Attention Module (MCAM), which fuses local features and global representations. Specifically, we propose three modules, convolutional neural network(CNN) module, convolutional channel attention module, and transformer module. Secondly, the convolutional channel attention module and CNN module are combined with transformer module to form the MCAM. Thirdly, we stack multiple MCAM to obtain our VeinCnnformer for vein recognition. Finally, the experimental results on three public vein databases show that our VeinCnnformer outperforms the existing vein recognition approaches and achieves the state of the art recognition results.

Federated learning is a novel distributed machine learning framework based on data privacy protection. In practical applications, there are often significant differences in data distribution among different clients in federated learning, which can lead to a decrease in the performance of federated learning models. This phenomenon is caused by the problem of independent and identically distributed data. To address the issue of independent and identically distributed data, this paper proposes a novel federated learning algorithm called FedCOA (Federated Cluster Optimization Algorithm). The FedCOA algorithm iteratively partitions the clients into different client clusters based on the similarity of their data distributions. Clients with similar data distributions within a cluster collaborate with each other, there by enhancing model generalization and reducing overfitting. In the simulation experiments, under the same dataset, the FedCOA algorithm achieves an approximately 10% improvement in accuracy compared to the FedAvg algorithm, and2% improvement compared to the CFL algorithm. Additionally, this paper conducts comparative experiments under different degrees of Non-Independent and Identically Distributed data distribution. The experimental results demonstrate that FedCOA can enhance model accuracy while effectively addressing issues such as the decrease in model accuracy and the inability of the model to converge caused by Non-Independent and Identically Distributed data distribution.

We examined young individuals' physiological responses and subjective evaluations toward different signpost design schemes across various conditions. The experimental methodology consisted of physiological measurements and subjective assessments to quantify the young participants' physiological indicators and preference levels toward various signpost design schemes. The experimental results showed significant variations in eye movement indicators and preference levels among the participants across different experimental conditions. In contrast, no variations were observed in electrodermal activity (EDA) and electrocardiogram (ECG) indicators. Notably, eye movement indicators demonstrated increased sensitivity to signpost design in the visual evaluation mode, influenced by the shape and lighting environment. However, preference levels were affected by color and pattern. These research findings offer valuable insights for signpost design and the evaluation of outdoor products.

At present, image processing has widely utilized in tremendous downstream applications including robotics, detection sensors and industrial products. Existing processing methods have achieved considerable accuracy towards complex images. However, existing classification concentrates on the convolution operation and ignore extraordinary information in feature space. In this work, we propose a novel classification based on the extracted features of images and utilized most advanced transformer model to identify the information from diverse images. Initially, the model extracts the features from the feature space and identify the feature information. Subsequently, the transformer model will train the classification mechanism through recognizing the feature information with the target labels. Finally, the trained model can recognize the different objectives in real-world images. From our extensive simulation and comparison analysis, we can conclude that proposed model can achieve the target classification with acceptable accuracy and reasonable training costs.

This paper focuses on high-precision 3D object detector. In two-stage 3D object detection algorithms based on point-voxel method, the concatenation of multiple types and scales of features is often limited to simple concatenation, which fails to capture rich semantic information. To better utilize multi-scale voxel features, raw point feature and BEV feature, we introduce a method called hybrid cross feature fusion module (HCFF). This method leverages a gating module to fuse raw point feature, multi-scale voxel features, and BEV feature, resulting in more comprehensive semantic feature information, which improves the accuracy of object detection in all classes in KITTI. It is especially good at smaller classes. In the end, we enhance the original PV-RCNN and propose a higher-precision point-voxel based 3Dobject detection algorithm. The results demonstrate that our HCFF-Det3D network outperforms PV-RCNN in terms of accuracy.

Extracting entity relations from unstructured text is an important step in constructing a knowledge graph, but most current methods are ineffective in dealing with the complex problem of overlapping entities. In this paper, we propose a novel entity relation joint extraction model. It extracts subjects through pointer annotation, fuses the extracted subjects with the sentence vector, then inputs them into an attention layer (Attention Mechanism Based on Relative Position Embedding, AMBRPE) to enhance feature expression ability. Under predefined relation conditions, the model extracts objects corresponding to the extracted subjects to generate relation triplets. And the model can effectively solve the problem of overlapping entities through hierarchical pointer annotation. In addition, we introduce an Adversarial Training Component (ATC) into the model, which generates adversarial samples for training and acts as a text data augmentation method to improve model generalization ability. On the public datasets NYT and WebNLG, we conducted extensive experiments and the results show that our model outperforms the cascaded binary tagging framework (CasRel) by 2.1 and 0.9 percentage points, respectively. Moreover, the effectiveness of the proposed model is verified through ablation experiments.

The traditional TextRank algorithm is based on a graph model, and the extracted abstract generated by it does not deviate from the source text, but there may be issues such as polysemy. The BERT model can capture contextual semantics well. It can directly convert sentences in text into vector form, which can better solve the problem of polysemy. BiLSTM can obtain further information, IDCNN can better handle overfitting problems compared to traditional convolutional neural networks, and CRF can better solve the problem of sentence coherence. The model proposed in this article has been compared with other extraction based text summarization methods on the TTNews dataset, and experiments have shown that the method proposed in this article can obtain meaningful text summaries with significant optimization.

Based on the development requirements of modern vocational education and the development trend of information-based teaching, the purpose of the paper is to build a three-dimensional and diversified teaching evaluation method by introducing data mining technology, use appropriate data mining algorithms to summarize the knowledge as well as master rules, predict the learning status of students, timely carry out early warning and teaching intervention, and explore a teaching evaluation and academic early warning method suitable for vocational education courses through practical classroom practice so as to improve curriculum mastery level of the students and optimize the quality of personnel training technical skills.

In the absence of instance-level supervision information in weakly supervised object detection, there are currently two primary challenges. The first challenge is the difficulty in clearly distinguishing multiple target instances, leading to issues of missed detection and confusion. Another issue is that during the training process, it is easy to be trapped in local optimal solutions, making it difficult to comprehensively detect target instances. To address the above problems, a weakly supervised object detection algorithm is proposed that combines attention mechanism and multi-instance mining. The algorithm adds incorporates an attention module enabling the detector to detect the full view of the target and alleviate local optima issues. Additionally, a pseudo-Instance mining algorithm is designed specifically for multi-instance detection, to discover instances overlooked by the detector, reducing instance omissions and thereby enhancing detection performance. The algorithm was evaluated using the PASCAL VOC2007 and VOC2012 datasets, demonstrating that it can detect more instances compared to other weakly supervised algorithms, effectively enhancing weakly supervised detection accuracy.

Lysine crotonylation (Kcr) is a form of post-translational modification that can be widely found in both histones and non-histone proteins. Precise detection of Kcr is essential for a better understanding of their functional mechanisms and drug development studies. Computational algorithms can predict Kcr and provide reliable data support for experimental methods. In this study, an effective new neural network-based computational framework for Kcr prediction (We-Kcr) was designed on a large amount of protein sequence data. After validation assessment and method comparison, the results showed that the prediction model exhibits higher accuracy in predicting Kcr sites. The research in this work provides biologists to explore the functional and biological mechanisms of protein modification. The materials can be downloaded from the following website: https://github.com/lihh225/We-Kcr.

Due to the strong reflective properties of the spacecraft surface coatings, there are significant challenges in processing images from outer space. Furthermore, the volume of data for image feature processing and matching is immense, and existing algorithms are insufficient for aerospace system applications. Therefore, this paper proposes a three-dimensional pose measurement algorithm based on binocular vision. The binocular vision-based three-dimensional pose measurement system consists of four main components: camera calibration, camera rectification, stereo matching, and pose determination. Traditional image processing algorithms are employed for satellite image processing. Camera calibration is performed using M software, and the calibration results are further optimized. Due to real-time requirements, an improved SIFT algorithm is used to detect local features in the images, extract feature points, and perform feature point matching under epipolar constraints. Experimental results demonstrate that this algorithm can achieve accurate and fast three dimensional pose measurement.

Multivariate time series forecasting has a large number of applications in the fields of meteorology, investment, and energy. Transformer-based prediction models have achieved good performance, but their computation of global correlation is responsible for a high degree, and they cannot effectively model local features like convolutional neural network(CNN) structures. As for the multi-scale dependency aspect, current research focuses on learning at fixed scales, including the dependency in time and frequency domains, and we hope that the model can automatically learn the appropriate scale features. In this paper, we propose a prediction model based on fractional Fourier transform, which decomposes the sequence, projects the data to different fractional domains using fractional Fourier transform, learns different global features, and can effectively obtain unconstrained scale dependence, and realizes cross-scale feature extraction, and fully learns the potential features in the sequence, and for the local information we utilize convolution to extract them, and the proposed model is called Fractional Fourier Convolutional Network (FRCNet). We have conducted extensive experiments on several publicly available datasets, including electricity, diseases, and exchange rates. The experimental results show strong competitiveness in multivariate time series forecasting.

To address the problem of large parameter and computation complexity of feature extraction network in correlation filter based target tracking algorithm, a lightweight correlation filter target tracking algorithm Iterative Enhanced Ghost Network （IEGN）IEGN is proposed. First, GhostNet is used as the backbone feature extraction network, and iterative attention feature fusion is used to increase the feature's perception ability of context. Second, the extracted deep features and handcrafted features are windowed and interpolated, and then convolved with the current filter in the Fourier domain for localization calculation. Finally, the conjugate gradient algorithm is used to optimize the loss function of the sum of response error and penalty term, and realize the filter update. Through experiments, it is concluded that the tracking speed is increased by 25% when there is GPU, and 60% when there is no GPU. The test tracking obtains 0.530 tracking accuracy, 0.362 robustness and 0.244EAO, which has the best tracking effect among the mainstream correlation filter algorithms in comparison. Compared with ResNet-50 network and VGG-16 network, this method reduces the feature parameter extraction amount by 80% and 96.4%, respectively.

The problem of separating mixed audio signals into independent speech signals is referred to as speech separation problem, also known as the cocktail party problem. We propose a novel method that utilizes TasNets to separate mixed speech signals. Similar to Dual-Path RNN, our innovation lies in the use of parallel recurrent neural networks within the block sequence, with an additional recurrent neural network added. We utilize fast attention to achieve global attention, aggregating context-aware information and achieving parallelism. We use the Swish activation function in the separation module. Our approach consistently outperforms GALR in the benchmark WSJ0-2mix task, with an absolute improvement of 0.7dB in SI-SNRi, for models of a similar size to GALR. Experimental results demonstrate the effectiveness of our network compared to previous methods.

This article proposes an insulator detection method based on optimized YOLOv5s. By optimizing and improving the structure of the YOLOv5s model, the accuracy and efficiency of insulator detection have been improved. At the same time, this article also proposes a data augmentation method for insulator detection, which increases the diversity of the training dataset. The experimental results show that the detection accuracy and recall rate of the proposed methodare0.95 and 0.97 respectively when the IOU threshold is 0.5, and the F1 score is 0.951. Compared to traditional networks in terms of accuracy and efficiency. It can meet the requirements of robustness and accuracy for insulator detection in transmission lines.

The rapid development of the smart home industry is accompanied by various challenges such as standardization, equipment compatibility, communication stability, network security, and device management. To address these issues, we propose a smart home system based on digital twin technology that includes wireless communication, intelligent algorithms, sensing, and security. Our system can simulate and predict the behavior of the real system, providing an efficient way to optimize performance. Using this framework, we design a smart lighting system that can intelligently adjust the lighting level and color temperature according to the user’s preferences and the surrounding environment. Overall, the proposed system can help enhance device interoperability and user experience in the smart home.

License plate location is one of the key technologies of license plate recognition. The location and detection of license plate is greatly affected by driving direction, weather anomaly, lighting and other factors, and often has license plate tilt, occlusion and other situations. Photos taken in these complex scenes will have adverse effects on the location detection of key information in the image. Aiming at the above problems, this paper proposes an improved license plate location technology of Yolov5. This method adds DAFS dynamic anchor frame to the backbone network of Yolov5 to effectively and quickly detect the license plate position. Compared with other methods in complex scenes, the experimental results show that the positioning and detection accuracy of the proposed method is improved by 1.08%.

Physical attacks on object detection models can be carried out using adversarial patches generated through deep learning. If the object is applied with the adversarial patch, the detector cannot find the object in the image. However, the patch generated by the traditional method often looks obvious and strange, making it very conspicuous in the physical world and hard to attack the object detection model covertly. In this paper, we propose a new method that combines style transfer with traditional patch generation to enhance the semantic content of images. Using style transfer, we can change an image's style while preserving its original character. We try different style transfer algorithms and compare each patch's attack performance, finding that using the convolutional neural networks based algorithm on the patch can make the patch's attack performance reserve the most.

Hashing has recently received excellent performance for similarity search on large-scale data set. In many scenarios, most image contains multi-instance features, and how to hash multi-instance image is an important task. Although some researchers focus on this problem and put forward solutions, they usually simply combine with existing hashing methods, which are not efficient methods for multi-instance hashing. To address this issue, we propose an Efficient Multi-Instance Hashing (EMIH) method. EMIH first construct bag features by multi-instance feature mapping. To improve the quality of hash codes, EMIH are both consider semantic correlation and local similarity preserving between bag features. Different from existing methods that relaxing the binary constraints, EMIH learn discrete codes directly by an alternate algorithm. Experimental results on two benchmark datasets demonstrate the superiority of EMIH as compared to related multi-instance hashing methods.

The Joint Video Exploration Team published the Versatile Video Coding standard in 2020. The standard improves the efficiency of coding for 360-degree panoramic video. However, it also causes an increase in coding complexity, especially in the selection of the intra prediction mode. Therefore, in order to reduce the time required for intra prediction mode decision in the VVC standard, this paper proposes a fast decision algorithm for intra prediction modes from the perspective of reducing the number of candidate modes in the intra prediction modes by exploiting the relationship between the gradient direction and the angular prediction modes of the Sobel operator. Experimental results show that this algorithm can save an average of 6.27% of coding time compared with VTM11.0, while causing an average increment of BDBR of only 0.198% and an average reduction of BDPSNR of only 0.009dB.

In recent years, the health problems of the aged persons have been paid more and more attention in China. Falls are one of the main reasons affecting the health and safety of the aged, especially accidental injuries caused by falls of the aged persons have attracted much attention. In this paper, the STM32F103C8T6 chip and the MPU6050 sensor are used to collect the posture information of the human body in the normal state and the falling state, and the Fast Fourier Transform was performed on the collected posture data to obtain the frequency data of human behavior. The frequency information is used as the dataset for the random forest model. Five-fold cross-validation strategy was used to increase model performance. Besides, different filtering algorithms were utilized and compared. The model accuracy reached94%, which demonstrated that the proposed algorithm in this paper is promising for better human fall detection problem.

Aiming at the problem that the existing UAV group target detection algorithm does not consider the connection between group individuals and is prone to miss detection of UAV, this paper proposes a group target detection algorithm for UAV group based on group structure. Based on the group structure information, the Group Structure Relation module is constructed to improve the detection and positioning ability of UAV group targets. Experiments were carried out on the constructed dataset. The experimental results show that the mAP of the proposed UAV group target detection algorithm reaches 91.3%, which is about 2% higher than that of the original YOLOv5 algorithm, effectively improving the detection accuracy of UAV group members. At the same time, the detection speed reaches 65 frames per second, realizing the real-time detection of UAV group targets.

Cracks in infrastructure engineering projects have a significant impact on the performance and functionality of roads, bridges, and buildings. Traditional crack detection methods are time-consuming and subjective, thus calling for faster and more accurate detection methods. This paper proposes a novel algorithm called the Multi-Layer Supervised Network, which combines the U-Net model and deformable convolution to address the challenges in crack detection. Our method utilizes the encoder-decoder architecture of the U-Net model to extract crack features and addresses the issues of gradient vanishing and overfitting through a multi-layer supervised mechanism. Meanwhile, the deformable convolution is introduced to be able to deal with different scales and angles of cracks adaptively and improve the feature expression capability. The experimental results demonstrate that our proposed method can accurately detect and segment cracks in images with good generalization ability. Compared to traditional image classification and object detection methods, our approach can extract more accurate crack information, providing valuable insights for the evaluation and maintenance of infrastructure engineering projects.

To enable the deployment of target detection algorithms for mobile device, we propose a lightweight improved algorithm based on YOLOv5s. Firstly, we propose a four-scale prediction network to enhance detection of small targets by adding a 160×160 prediction head to the Head layer of the network. Secondly, MobileNetV3s is chosen to replaceCSPDarkNet53 as the backbone network, and CA (Coordinate Attention) is introduced into the MobileNetV3s. We use Ghost Module to replace the traditional convolution in the Neck part. Finally, We use SIoU_Loss to replace CIoU_Loss. The experiments’ results show that the Flops of this paper has decreased by 41.8% and the model size has decreasedby65.1% relative to the YOLOv5s before the improvement, and the mAP@.5 has reached 93.3%, providing theoretical support for the deployment of vehicle-pedestrian detection algorithm in mobile device.

In recent years, campus security accidents frequently occur. The establishment of a set of campus security systems based on artificial intelligence can allow schools, parents, and relevant government departments to make timely warnings and respond to emergencies, effectively protecting the safety of children, which has become an important task in the construction of the current campus security system. This paper designs a smart campus security system based on computer vision and Internet of Things communication technology. The system is mainly analyzed by a vision algorithm, and the campus crowd detection system is designed. The YOLO algorithm based on the DarkNet53 network is mainly used to separate pedestrians from the background for pedestrian detection. Through perspective transformation technology, the center coordinate of the image pedestrian is converted into a three-dimensional coordinate, and the crowd detection and the crowd danger action detection alarm are realized.

The artificial intelligence technology has been widely applied in agricultural production. In particular, scientists have designed special robots in different agricultural scenarios, such as Intelligent Pesticide Spraying Drone, Farm Weeding Robot, etc. Here we developed a fruit picking robot, where the accurate identification and discernment of fruits hold a significant role. The deep learning technique is applied to conduct fruit detection and recognition. This technique adopts YOLOv5s as the base network. The network has been trained with a collection of six types of fruit pictures. The empirical findings have unveiled the effectiveness of this fruit detection and recognition technique on diverse test image datasets from different kinds of fruits.

In response to the problem of low efficiency and accuracy in detecting welding defects on the surface of coal mining machine drums, a drum welding defect automatic detection method based on the improved YOLOv7 algorithm is proposed. By utilizing the characteristic of long-distance fusion of multiple feature information in the Transformer model, the global context CoT model is introduced into the YOLOv7 algorithm network, and experimental verification is conducted. The results show that this detection method has high detection accuracy in welding defect detection and can eliminate the image redundancy problem caused by the introduction of attention mechanism. It can be applied to the quality inspection of coal mining machine drum welding products.

To solve the problem that the traditional general algorithm can not detect small targets well, this paper adds a new feature fusion branch and introduces an attention mechanism on the basis of yolov4 algorithm, which avoids feature aliasing caused by simple fusion and enhances the expression ability of the network for small targets. Compared to the original network, mAP improved by 2.2%. On this basis, considering that the low confidence prediction box also contains the prediction information of small targets, but the traditional method directly rejects it rudely, resulting in missing detection. In this paper, the K-means clustering algorithm is used to analyze the size distribution characteristics of the targets in the data set, then the image is divided into regions, the region containing the low confidence prediction frame is enlarged by super resolution, refined detection, the results are mapped back to the original location, and the results are uniformly processed by NMS. In the end, compared with the original network, mAP improves by 4.4%, and the network's recall rate of small targets has been significantly improved by nearly 8%, which can effectively improve the performance of small target detection.

The use of near-infrared spectroscopy enables rapid and efficient detection of tobacco leaf chemical components. However, changes in environmental and climatic factors can affect the accuracy of the originally established prediction model. To reduce the workload involved in re-sampling and modeling, a more stable prediction model was established using transfer learning, transferring valuable information from fragmented leaf spectra to enhance the stability of the whole-leaf spectral measurement model. New prediction models were developed for total sugars, reducing sugars, total alkaloids, potassium, and chlorine contents. Experimental results show that the mathematical model established through transfer learning provides highly consistent predictions with results obtained through continuous flow methods. It is more stable than existing models and can be directly applied for model transfer without the need for re-sampling. It can be used for the analysis of tobacco leaf chemical components in the tobacco production process, meeting the requirements for rapid detection of routine chemical components in tobacco leaves. The model exhibits high accuracy in predicting new samples and can meet the demands of tobacco production companies.

In order to solve the problems of large number of deep model parameters and poor real-time performance in the existing vehicle detection methods, this paper proposes a vehicle detection method based on Taylor expansion model pruning algorithm. VGG16 is the main trunk network, the Taylor expansion method is is used for filter pruning, and the Lagrange algorithm is combined. Experiments show that the vehicle detection method proposed in this paper can reduce the complexity of the model, reduce the number of model parameters by 90%, and shorten the detection time without affecting the detection accuracy, which has certain application and promotion value.

The essence of weak and small target detection is to achieve separation between the target and the background. However, for extremely low contrast (<0.1) and extremely low signal-to-noise ratio (<=3dB, most of which are less than 0dB) targets in strong backgrounds, the intensity difference between the target and the background is very small. However, in strong backgrounds, the fluctuation of the target submerges it and makes it difficult to directly separate it. Therefore, in order to achieve target detection in such extreme scenarios, it is necessary to construct feature descriptions that can effectively separate them. Based on this idea, this article summarizes the target detection task as a local feature difference maximization model suitable for all spatial target detection, and uses the local maximum Pearson correlation coefficient as the feature extraction equation to calculate the correlation between the two patches. Based on the small correlation between the target and the local background, and the high correlation between the background and the background, the separation of the target and the background is completed. Then using a constant local signal-to-noise ratio feature extraction equation to enhance the Pearson correlation results. A large number of experimental results show that the model and algorithm proposed in this paper can effectively detect targets with extremely low signal-to-noise ratios in strong backgrounds.

Polarization imaging is capable of effectively showcasing the polarization properties of objects while also mitigating the impact of strong light and enhancing the visibility of weak light. As a result, it compensates for the limitations of intensity imaging in low-light environments. To fully utilize the advantages of intensity and polarization images, this paper proposes a method that combines NSCT decomposition and an improved dual-channel PCNN model. The method enhances the traditional PCNN model by extending the input to dual channels, employing adaptively calculated parameters, and applying adaptive linking weights to the high- and low-frequency images obtained from NSCT decomposition. The highfrequency image utilizes the MSMG operator, while the low-frequency image employs a weight map controlled by multiple parameters. This method is utilized for fusion processing of intensity and polarization images.By conducting comparative experiments with six commonly used algorithms, the results demonstrate the superior performance of this method in terms of preserving texture details, providing high-quality images, achieving information-rich fusion, generating fused images with rich grayscale levels, and maintaining structural similarity. In summary, the proposed method exhibits significant advantages in fusing intensity and polarization images.

In the context of the rapid development and widespread application of remote sensing technology, small object detection has become a prominent research focus. Despite the extensive use of the YOLOv5 network in the field of object detection, its performance in detecting small objects, especially in remote sensing images, remains unsatisfactory. Particularly, detecting and recognizing small objects, such as aircraft, pose greater challenges. The reasons for this include the small size of the targets, low contrast between targets and backgrounds, and the lack of comprehensive publicly available datasets. To address these issues, this study constructed a dataset of remote sensing images containing small aircraft targets, which facilitates the network in capturing fine-grained features and improving detection performance, thus compensating for the shortcomings of existing publicly available datasets. Based on the YOLOv5network model, this study proposed the following optimization measures: (1) To tackle the issue of small target sizes, the model structure was simplified to make the feature extraction network more suitable for small objects and to reduce the number of model parameters. (2) In response to the deficiencies in the original model's fusion method, a bidirectional Feature Pyramid Network (BiFPN) was introduced to enhance multi-level feature fusion capability. (3) To reduce the computational complexity of the model, reasonable anchor boxes were designed to enable the model to accurately focus on crucial information during the detection process. Experimental results demonstrate that the proposed algorithm improves the detection accuracy and speed of small aircraft targets in remote sensing images. On our custom dataset, the method achieves excellent results in terms of precision, computational efficiency, and parameter count.

This thesis investigates the innovative application of virtual reality technology in the digital display of intangible cultural heritage. Intangible cultural heritage is an important part of human culture, and with the continuous development of digital technology, the protection and display of traditional intangible cultural heritage is facing new challenges. Virtual reality technology as an emerging technical means, its digital display is of great significance for the protection and inheritance of intangible cultural heritage, and its immersive and interactive features can bring a new experience for the display of intangible cultural heritage. This paper firstly introduces the basic principles and characteristics of virtual reality technology, and then explores the application methods and effects of virtual reality technology in the digital display of intangible cultural heritage. It is found that virtual reality technology can enhance the audience's cognition and understanding of intangible cultural heritage by simulating realistic scenes and providing interactive experiences. Finally, this paper proposes future directions and challenges for the development of virtual reality technology in the digital display of ICH, as well as suggestions for further research. This study is of great significance in promoting the protection and inheritance of intangible cultural heritage and the application of virtual reality technology in the cultural field.

Image watermarking technique involves three important attributes of robustness, imperceptibility and watermarking capacity, which constrain each other and need to be weighed in designing watermarking schemes. Improving watermarking imperceptibility, watermarking capacity and robustness has been a challenge for image watermarking algorithms. In this paper, the algorithm utilizes the magnitudes of fast and accurate polar harmonic Fourier moments (FAPHFMs) as the watermark carrier, which improves the robustness of the algorithm. Using the multiplicative embedding method, the watermark embedding strength can be flexibly adjusted according to the size of the original coefficients to achieve better robustness and imperceptibility. Experimental verification confirms that the algorithm outperforms others in terms of robustness and imperceptibility.

The wheel hub is an important component of automotive components, playing a supporting and braking role, and its quality directly affects the stability, safety, and comfort of the vehicle during driving. With the development of the automotive industry, enterprises have increasingly high requirements for the quality of wheel hubs, so the detection of wheel hub quality has practical significance. To meet this requirement, this article proposes a processing algorithm for determining the quality of wheel hubs through harmonic analysis. This algorithm uses the 12 point coordinate method to equidistant 12 pairs of measured values from the measured wheel hub data. The 12 pairs of measured values are substituted into the harmonic analysis formula to obtain the Fourier factor. Based on the obtained Fourier factor, the amplitude and phase of each harmonic are calculated, and then each harmonic of the wheel hub is plotted. The jumping amount of each harmonic obtained is compared with the given judgment threshold to determine whether the wheel hub quality is qualified. The results show that the algorithm can obtain the sixth harmonic of the wheel hub at the highest, and the conclusion can be drawn by calculating the variance, which is more accurate than the commonly used algorithms before.

Density Peak Clustering (DPC) is a density-based clustering algorithm with good overall performance. However, clustering on data sets with large density differences or irregular distribution is prone to "false peak" or cluster loss, resulting in data point-associated assignment errors. To solve the above problems, this paper proposes a mutual nearest neighbor search algorithm based on dynamic neighborhood, which uses the dynamic neighborhood to select the adaptive K value, and combines the mutual neighborhood matrix distribution data points to detect the variable density cluster and the irregular distribution cluster with the least parameters. After experiments on 11 synthetic datasets and 3 real datasets, and comparison with DPC, K-means, and Extreme_Cluster method, the results show that the proposed method has strong robustness for clusters with different density distributions and special shapes.

Content Aiming at the problem that iris images are easily affected by eyelid and eyelash noise, which leads to low positioning accuracy and poor stability, an iris positioning network based on YOLOv4 model is proposed. Firstly, the backbone feature extraction network combining ECA attention mechanism and Densenet121 is introduced to enhance the feature extraction capability of the model. Secondly, the SPP module is improved and replaced with a deep detachable void space convolution pool pyramid to promote the information interaction between low-level location information and high-level semantic features. Finally, a self-organizing particle swarm optimization algorithm is designed to obtain the prior frame size that can characterize the overall features of iris data set. The proposed method was tested on the CIASIA-IrisV4 iris dataset, and the average accuracy of the improved model was 99.70%, and the positioning accuracy of the improved model was 95.89% in the case of light reflection generated by glasses. Compared with other iris positioning models, the accuracy and robustness were higher, which had obvious advantages.

In the context of vehicle operation, detecting and identifying abnormal driver behaviors is challenging due to the complex in-car environment, changing lighting, and varied driver postures. Our approach, utilizing supervised contrastive learning, categorizes driver behaviors as normal or abnormal. We employ depth images from the driver's front and above to address environmental complexities and improve accuracy. Our enhanced 3D-MobileNetV2architecture achieves impressive results on the DAD dataset test set, with a 94.18% accuracy rate and a 0.962AUC, validating the effectiveness of our method in driver anomaly detection.

The knowledge graph-based question answering system aims to find the connection between the internal knowledge behind the problem and the known knowledge base triplet. This paper proposes a framework for knowledge graph question answering (ZDNN-KGQA), which divides intelligent question answering based on knowledge graphs into four steps: named entity recognition, entity disambiguation, attribute classification, and answer selection. In the named entity recognition step, this paper proposes a BERT-BiLSTM-CRF model that combines an attention mechanism to identify entities in question sentences. In the entity disambiguation step, this paper utilizes the BERT-Softmax model to extract the accurate entities from the candidate entity set. In the attribute classification step, it is converted into a problem of text similarity. This paper proposes the use of BERT-Softmax and Word2vec models to extract attributes intext problems, effectively improving entity recognition performance and question semantic parsing performance. The average F1 value of this system on the test dataset provided by the NLPCC-ICCPOL 2016 KBQA task is 0.8974, which is close to the optimal performance level.

Many studies show that convolutional neural networks can improve performance after embedding attention mechanism. However, the existing related research either develops more complex attention modules in pursuit of ultimate performance improvement; Or the performance improvement is not obvious in pursuit of ultra-lightweight In order to balance the contradiction between the performance and complexity of attention mechanism, this paper proposes a simple convolution pyramid attention module (CPA). Firstly, the module constructs a convolution pyramid which can be used for multi-scale feature extraction by stacking two-dimensional convolution with convolution kernel size of 1, and then counts the global information of feature mapping in the channel dimension. Then, the feature fusion is carried out with the exclusion block, and the output weight vector is multiplied with the original feature mapping in the channel dimension, thus realizing the adaptive correction of feature mapping in the channel dimension. Finally, the performance superiority of the proposed attention mechanism is proved by image classification and target detection tasks, and then its structural rationality is proved by ablation experiments.

For the UAV inspection process to detect the target size is small, the distribution of dense and serious obstruction and other reasons leading to the target detection difficult problem, a small target detection algorithm for UAV inspection scene based on SPD-Conv is proposed. SPD-Conv is a convolutional layer of a deep learning model, which contains a space-to-depth layer and a non-step-size convolutional layer, and SPD-Conv is a deep learning model. Conv replaces each step-spanning convolutional and pooling layer to reduce the loss of detail characteristics during downsampling and improve network's performance in recognising small targets. By adding a 160×160 small-target detection head and deleting large-target detection head, SA attention mechanism is added before each detection head to reduce interference and improve network's ability to extract features from small targets. Experimental results on VisDrone2019 dataset show that the improved model has improved misdetection and false detection phenomena. The detection accuracy of improved model is 44%, which is 6.6 percent higher than that of YOLOv5s model, and speed reaches 84.03 FPS. The improved model can fulfil real-time detection requirements and detection accuracy is improved.

Aiming at the problem that traditional visible light imaging systems are easily affected by environmental factors and cannot accurately describe the scene information, a gradient residual generative adversarial network for polarization image fusion is proposed in this paper. In the generator network structure, dense blocks are added to the encoder part to retain more features, the gradient residual module is constructed for fusing the feature maps to enhance the texture and detail features of the image, and the multi-scale weighted structural similarity loss function and gradient loss function are designed to improve the network performance. The experimental results show that the method in this paper obtains fused images with richer texture structure and more in line with the visual sense perception of the human eye, and at the same time has the optimal objective evaluation index.

The knowledge graph in the knowledge graph is sparse due to missing relationships and incomplete triples, which promotes research on knowledge graph completion (relationship prediction). However, the purpose of knowledge graph reasoning technology is to infer unknown knowledge or identify wrong knowledge based on the knowledge learned in the knowledge graph to improve the knowledge graph. The graph attention network (GAT) can improve the input constraints of RNN and convolution and introduce a self-attention architecture to perform classification of graph-structured data. However, the model still has two limitations: applying the model in the knowledge graph will ignore the relationship (edge) features; Currently popular models can only use the k-th hop output for multi-hop embedding learning, resulting in the loss of a lot of early embedding information (such as one hop) during the graph attention process. This paper proposes a new method to improve the quality of the model by embedding relational features.

This study employed a training dataset comprising 10 distinct vehicle models for prediction purposes. Among a plethora of models, InceptionV3, MobileNet, and DenseNet were selected due to their relatively high accuracy rates, and they were utilized to enhance the algorithm. Comparative analyses were conducted with the models prior to improvement. The primary approach in this study involved data augmentation, which included rotating images by specific angles, scaling, proportionally shifting images horizontally, and performing horizontal flips on the images within the dataset. Various performance metrics, namely Accuracy, Precision, Recall were calculated using specific formulae. The predictive outcomes for the classification problem were summarized, and specific metric data for different vehicle models were observed to determine the most suitable model for our dataset. A comparative analysis of the predictive accuracy of different network architectures for different types of vehicle models has been determined.

Loose-leaf textbook is a new type of textbook in the new era of education environment, which has better interactivity and flexibility of use than traditional textbooks. In order to improve effect analysis of teaching with loose-leaf textbooks, a multi-dimensional evaluation method of teaching with loose-leaf textbooks based on the improved K-modes algorithm is designed. In the process, the multi-dimensional evaluation system of teaching using loose-leaf textbooks is constructed by combining objective and subjective directions, and the corresponding index value calculation method is designed, and then the data processing of multi-dimensional evaluation based on the improved K-modes algorithm is optimised by introducing the error sum of squares and the co-presence rate to optimise the clustering process, so as to obtain a complete multi-dimensional evaluation method. Results show that the research method maintains the accuracy rate above81.2% when the number of sample points increases to 1200 when performing data classification accuracy analysis; when generating the multidimensional evaluation results, the research method generates realistic evaluation results for 8indicators. It shows that the research method can effectively carry out the multidimensional evaluation of teaching using loose-leaf textbooks, and can provide certain technical references to the education industry.

Ophthalmic diseases are prevalent worldwide, encompassing a diverse range of conditions such as cataracts, glaucoma, and retinal diseases. The World Health Organization (WHO) reports that more than 2.2 billion individuals globally suffer from visual impairments, with nearly half of these cases being preventable through early diagnosis and timely treatment. However, ophthalmologists encounter multiple challenges in clinical diagnosis, including the complexity of cases, the absence of clear early symptoms leading to delayed detection and intervention, and a scarcity of medical professionals in remote and developing regions. To address these issues, Artificial Intelligence (AI) has emerged as a potential solution. This research aims to explore the application of AI in the classification of three common eye diseases: cataracts, retinal diseases, and glaucoma. Specifically, we compare the performance of six Convolutional Neural Network (CNN) models and Swin-Transformer models using the same dataset. The objectives of this research are twofold: (1) to demonstrate the potential of AI in diagnosing eye diseases, and (2) to determine whether Transformer models outperform CNN models in terms of eye disease classification. To achieve these objectives, we have compiled a comprehensive dataset including images of cataracts, retinal diseases, and glaucoma. Through transfer learning, we trained and evaluated models on this dataset. This study offers valuable insights into the application of AI in diagnosing eye diseases, potentially paving the way for the development of AI tools that can support medical professionals in accurate diagnosis and treatment. Ultimately, the integration of AI in ophthalmology could enhance patient care and outcomes.

Aiming at the particle filtering algorithm's particle degradation and impoverishment easily lead to the decrease of filtering accuracy, which makes the target tracking accuracy and robustness decrease, a particle filtering algorithm based on the improved sparrow search algorithm (ISSA-PF) is proposed. Frog Leaping Algorithm (FLA) is used in the exploration phase of SSA to update the location of the discoverer. Self-adaptive adjustment of step size based on individual fitness values. Less adaptable individuals use larger step sizes to quickly jump out of local optima. The better adapted individuals use smaller step sizes to search the surrounding solution space more finely, thus improving the algorithm's optimization performance and search capability. Introducing a Levy flight strategy to simulate sparrow flight trajectories to perturb the optimal solution for each generation of the population. Expanding the algorithm's search range and accelerating convergence. The experimental results show that the algorithm proposed in this paper has higher target tracking accuracy as well as faster operation speed than the traditional Particle Swarm Particle Filtering algorithm (PSOPF), Gray Wolf Optimization Particle Filtering algorithm (GWO-PF), and Sparrow Search Particle Filtering algorithm (SSA-PF).