This PDF file contains the front matter associated with SPIE Proceedings Volume 12916, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

The crack of subway tunnel seriously affects the service life of the tunnel and endangers the safety of subway traffic. The efficiency of manual detection is low, this paper proposes a tunnel crack detection method based on image processing. The images of tunnel cracks collected inside the subway tunnel suffer from problems such as uneven lighting, low contrast, and severe noise, which affect the recognition of cracks. To address these issues, a multi-scale Retinex algorithm combined with linear stretching is firstly used to preprocess the images, effectively balancing the lighting. Secondly, an improved adaptive median filter algorithm is used to filter out image noise while effectively preserving the crack edges. Thirdly, the Scharr operator combined with the Otsu method is used to segment the filtered image, effectively separating the crack area. Fourth, the crack binary image without noise is obtained using connected domain filtering and morphological processing. Finally, the crack length, average width, maximum width, and area are calculated using the crack skeleton image. The research results show that the proposed algorithm can effectively identify crack areas, demonstrating its effectiveness.

With the continuous progress of science and technology, image processing technology is increasingly used in real life. Due to a series of advantages such as large amount of information, fast transmission speed, and long operating distance, image processing technology has become the main source of human information acquisition and an important means of utilizing information. In recent years, morphological image processing has become an important direction in image processing. Early image processing was mainly aimed at improving the quality of images, taking people as research objects, mainly to improve people's visual perception. In the process of image processing, various methods such as image enhancement, restoration, encoding, and compression are commonly used to process images. Many scholars at home and abroad are committed to the research of computer image processing, including image segmentation, image compression, image restoration, and other aspects. This paper introduces an image preprocessing method based on wavelet transform. In conventional preprocessing, we use smooth and sharp methods. For example, for Gaussian noise, we usually use linear filtering; When dealing with salt and pepper noise, median filtering is usually used, and better filtering results can be obtained. Then analyze the advantages of algorithms from the perspective of algorithm data, and have the courage to innovate to find algorithms to alleviate and eliminate imbalances. The research results obtained are expected to serve as a reference and supplement for the current research on image processing algorithms in China.

Publisher's Note: This paper, originally published on 20 October 2023, was withdrawn on 13 December 2023 per author request.

Tree barrier plays a vital role in transmission lines, and its existence directly affects the stable supply of electricity. Therefore, it is necessary to measure and inspect the tree barrier. For this purpose, this paper proposes a measurement method based on monocular image depth estimation. In this method, the LSD (Line Segment Detector) algorithm is used to detect the straight line, and the RANSAC (RANdom SAmple Consensus) algorithm is used to iteratively detect the vanishing point, and the vanishing point coordinates are successfully obtained. The camera calibration is carried out to determine the pixel coordinates of the object in the image and to obtain the internal and external parameters and rotation matrix of the camera. Through coordinate system transformation, the depth estimation of the target in the world coordinate system and the distance between any tree lines can be accurately calculated. Finally, the height of the camera is used as the reference height to complete the measurement of the tree barrier around the wire, and the experimental results show that the method has certain feasibility and the error is controlled within a reasonable range.

Binocular vision is an important branch of computer vision that processes images captured from two different cameras with different perspectives to obtain depth information of a scene. However, the stereo matching algorithm is complex and traditional platforms often struggle to meet real-time and accuracy requirements simultaneously. In this article, we implement a binocular vision system based on the idea of software-hardware co-design on the Zynq platform. According to the characteristics of the Zynq platform, the system's software and hardware functions are divided. The processing system (PS) is mainly responsible for flow control and binocular correction, while the programmable logic (PL) is responsible for parallel acceleration of the stereo matching algorithm. In the image acquisition part, we configure the format and output image timing using a designed camera acquisition module. In the stereo matching part, we design a semi-global stereo matching algorithm using Verilog language. Finally, we build a binocular vision system on the Zynq platform for testing. The experimental results show that the frame rate can reach 30fps when the input image resolution is 640×480, which satisfies the requirements of real-time and accuracy of the output disparity map.

This paper proposes a remote sensing image-based method for the extraction of fire trails. Firstly, by acquiring multispectral images of the forest in the study area and pre-processing the multispectral images. A vegetation correlation index is calculated, the vegetation correlation index is normalized, read into the vector range of the study area, and the vector range is cropped. Finally, the cropped area is extracted based on a first threshold, the extracted area is transformed, and the transformed area is filtered based on a second threshold to obtain the final fire-trail area. The invention is able to automate the pre-processing and extraction, using only one vegetation index, and combining with area thresholds, the process is more simplified.

Ground-based radar has been widely used for deformation monitoring and early warning of geological hazard potential areas. However, during long-term monitoring, ground-based radar images are vulnerable to human and environmental factors leading to severe decoherence. The study of ground-based radar image change detection can provide reference information for its long-term monitoring. Based on this, an unsupervised change detection method based on convolutional neural network (CNN) for ground-based radar images is proposed in this paper. First, the interference principle to extract change information is used for the first time for the change detection task, aiming to improve the accuracy of the initial extraction of change regions. Secondly, the fuzzy c-means clustering algorithm is used to obtain the pseudo-label matrix with categories, and the appropriate neighborhood image blocks with pseudo-labels are selected as training samples to train CNN. Finally, the change detection results of ground-based radar images are obtained using the trained CNN. Experiments were conducted using actual measurement data from ground-based radar in a monitoring task in a mining area in China and compared with other methods to verify the effectiveness of this paper's method and more accurate detection results.

Most of the existing imaging methods in MIMO radar require that the transmitting waveforms are orthogonal, the array architecture is regular and the sampling is uniform. In this paper, a computational imaging method for MIMO radar is proposed. The new imaging method is suitable for the case of nonideal orthogonal waveforms, irregular array architecture and nonuniform samplings. What is more, through the process of truncated singular value, the proposed method is robust even though the sensing matrix is ill-conditioned. The effectiveness of the proposed method is demonstrated by using the simulated data and the electromagnetic computation data.

Search and rescue operations play a vital role in saving lives in emergency situations. Knowing the exact latitude, longitude coordinates and directional angles can ensure effective and efficient operations when search and rescue personnel are operating in remote or hazardous environments. This paper presents the design and implementation of a search and rescue device based on magnetic compass. The device uses ESP8266 as the primary control module, combining GPS module for positioning and magnetic compass module for angle of acquisition. The information collected is corrected and transmitted to ESP8266 to provide feedback, which is displayed on the OLED screen to search and rescue personnel. The test results showed that the device performed well in all of the selected test scenarios, with an average accuracy of 0.61°for directional angle data and 1.17 meters for latitude and longitude coordinate data. This improved search and rescue device can improve search and rescue operations and increase the chances of successful rescues.

With the continuous improvement of living standards, passengers' requirements for ride comfort are increasing, and accurate human vibration comfort evaluation methods are very important to effectively improve ride comfort. At present, the human vibration comfort evaluation method mainly focuses on the analysis of the external vibration response signal of the human body, ignoring the response of the human body itself to the vibration stimulus, and when stimulated by the external vibration, each organ will produce a certain forced movement. Therefore, this paper attempts to explore the deep causes of human discomfort and proposes a method for evaluating human vibration comfort based on EEG signals.

This paper presents a method for multichannel filter banks on directed graph signals. The multichannel filter banks on directed graph are constructed with oversampling structure, and the analysis and synthesis filters are designed in the form of directed graph filters. In this paper, the spectral folding properties of the directed Hermitian Laplacian basic matrix on the directed bipartite graph are derived and analyzed, and the method on the undirected graph is extended to the directed graph by combining the signal representation on the directed graph, and the rotation parameters and the values of the matrix weights are also discussed. The results of the multiresolution analysis on the signals on different directed graphs verify the perfect reconstruction properties of the designed directed graph filter banks with the help of Chebyshev polynomial approximation algorithm.

In this paper, a novel full duplex Radio Frequency Passive Optical Network (RPON) system by using 8-ary pulse amplitude modulation (8-PAM) downlink and duobinary (DB) uplink signals is designed. We measure optical spectrum diagrams, eye diagrams, bit error rate (BER), and analyze the reception performance of the 10Gbit/s system before and after transmission. Results show that the downlink 8-PAM and uplink DB signals can improve bandwidth efficiency. Hence, the scheme is expected to be applied in the future high-speed full duplex access system.

Mobile signal strength affects the deployment of IoT devices, and its distribution is often measured using UAVs designed reasonable routes and equipped with measuring equipment. Traditional ant colony algorithms used in track planning can easily fall into local optimal solutions and is not suitable for large-scale tasks, and an improved ant colony algorithm is proposed for solving the problem. The characteristic of the improved algorithm is that in the process of updating pheromones, an information release function that changes with the number of iterations is introduced to avoid falling into local optimal solutions. Additionally, optimized the pheromone volatility factor to enhance global search capability. Theoretical analysis and simulation experiments demonstrate that compared with traditional ant colony algorithms, the improved algorithm can not only quickly escape local optima, but also has stronger global search capabilities and certain robustness.

Most of the existing polarization image based defogging methods use a priori and assumptions to recover images, and although these methods have made great progress, the a priori or assumptions are not always reliable in practical scenarios, which limits the performance of defogging methods. In this paper, we design a two branch network to learn image features; the defogging network uses the fusion module to adaptively assign weights to different path features to ensure the defogging effect while focusing on the haze region; the detail recovery network uses convolutional layers and smoothly expanding convolution to expand the perceptual field and fully obtain local and global feature information; finally, the features obtained from the detail recovery network and the defogging network are fused to improve the defogging Finally, the features obtained from detail recovery network and defogging network are fused to improve the defogging effect. The experimental results show that this method can reconstruct clear images in foggy environment with 2db improvement in PSNR and 0.016 improvement in SSIM, and the quality of reconstructed images is better than existing algorithms.

The paper presents a synchronization theory of series laser link consisted of three kinds of erbium-doped fiber ring lasers. Firstly, we present a synchronization equation of an erbium-doped fiber three-ring laser synchronizing with an erbiumdoped fiber dual-ring laser while a steady synchronization can realize between this two lasers when real parts of all eigenvalues are negative values. Secondly, we present another synchronization equation of the erbium-doped fiber dualring laser synchronizing with another erbium-doped fiber single-ring laser while a steady synchronization can realize between this two lasers when real parts of all eigenvalues are negative values. Moreover, a synchronization theory of the series laser link is proven by this two sets of synchronization equations. Our numerical result indicates that the series synchronization laser link are obtained in a chaotic dynamical state, a dual-cycle dynamical state, and a four-cycle dynamical state in three cases. We find that dynamics of the series laser synchronization link can be induced to all kinds of dynamical behaviors by adjusting the pump levels. The series laser link synchronization theory can be used as an important reference for laser secure communication, neural network, and optical network.

In practical industrial production, it is often difficult for machine learning to obtain sufficient image features due to the need for confidentiality or the scarcity of samples themselves. Therefore, this article conducts in-depth research in the field of Zero-shot Learning (ZSL). The key assignment of ZSL is how to infer latent semantic expertise between visible aspects of seen lessons and textual attribute features, so as to reap know-how switch to invisible classes. This article proposes an advanced algorithm in ZSL, which realizes the classification and recognition of unknown images by establishing a mapping relationship between local semantic attributes of text and images. The research in this article mainly includes the following aspects. Different from the traditional way of marking significant features manually, multiple different feature attributes are jointly used to guide the learning of global and local features of images. First, by using a text encoder and an image encoder, the text attribute words are encoded and embedded into the visual space, aligning the information of the two modalities in one dimension. Then, through self-attention mechanism, the semantic connection between attribute text and local visual information is established. Finally, through the classification module, the joint prediction attribute vector of global and local features is established, and the cosine similarity is used to predict the relative distribution of attributes between and within classes, thereby improving the generalization of prior knowledge in visible classes to invisible classes.

Waveform optimization technology based on phase encoding has become a key technology to improve the ability of radar to detect small targets. Designing different phase encoding models for different application scenarios and platforms can effectively improve the performance of radar in complex environments such as clutter and interference. Therefore, it’s very important to design an optimization algorithm with high orthogonality and fast convergence. This paper proposes an improved dynamic genetic algorithm to solve the optimization problem of Multi-Input Multi-Output radar phase encoding signal set. By improving the optimization model of the genetic algorithm, the diversity of the population is quantified to prevent the algorithm from converging prematurely. The improved dynamic genetic algorithm reduces the genetic probability of inferior individuals in the selection operation, then proposes to update the crossover probability in the crossover operation, and finally designs the mutation probability for individual gene points in the mutation operation, which solves the key problem of poor diversity in existing algorithms question. The simulation results show that the improved dynamic genetic algorithm improves the population diversity, optimizes the convergence speed of the algorithm, and the optimized phase encoding set has good performance, and the result is better than the existing improved genetic algorithm.

In order to solve the problems of long matching time and high mis-matching rate in the feature point matching process of traditional panoramic image stitching algorithms, this paper proposes a panoramic image stitching algorithm based on SIFT and RANSAC.SIFT's algorithm has been advanced to only extract features for overlapping areas (valid regions), thus lessening the computational effort of feature detection, better aligning image features, and enhancing alignment accuracy. On the ground of their corresponding degree, the RANSAC algorithm has been increasingly improved and thus allocated varying weights to distinct feature points, and the higher the matching degree, the higher the weight assigned to the points, thus more precisely discarding mismatched point pairs. A comparison of the number of mismatched points involving relevant parameters, like degree and time of matching, the SIFT and corresponding RANSAC algorithm-based panoramic image stitching algorithm are more excellent.

To address the issues of high measurement difficulty and poor measurement accuracy in traditional cattle body measurement methods, this study employed a multi-view stereo vision-based cattle body three-dimensional (3D) reconstruction and automatic body measurement method to conduct comparative research on a yellow cattle model and 47 dairy cows. An automatic data acquisition channel equipped with multi-angle stereo cameras was designed, and a cattle body keypoint recognition method based on the YOLOv7 object detection algorithm was used to achieve instant automatic multi-angle cattle body 3D point cloud data acquisition. To reduce the influence of point cloud noise caused by external factors such as lighting, this paper used statistical filtering and guided filtering methods for local point cloud denoising. To improve the accuracy and authenticity of the 3D reconstructed point cloud, a point cloud registration and fusion method based on the nearest iterative point was applied. Finally, polar coordinate transformation and curve fitting methods were used to measure data including body length, body diagonal length, body height, girth, chest circumference, and abdominal circumference. The experimental results show that the average relative error between the 6 sets of body measurement data and the actual measurements in the laboratory environment is 1.76%. In the breeding farm environment, the average relative error between the 6 sets of live cattle measurements and the actual measurements is 1.74%. This study can be applied to contactless cattle body measurement, which is beneficial for achieving precise cattle breeding.

Achieving high-quality output if large-scale SAR image stitching demands color balancing of the images to be stitched. However, adjacent SAR images may exhibit large differences in radiation intensity, particularly in the context of different-track SAR images. Furthermore, the diverse sizes and uneven spatial distribution of SAR images from different sources pose a challenge for color balancing in image stitching. To alleviate these issues, we propose a large-scale multi-track spaceborne SAR image dodging stitching scheme based on the working characteristics of spaceborne SAR. We leverage different algorithms for along-track and across-track SAR image dodging, based on the statistical characteristics of the overlap region and block processing, to ensure uniform luminance and contrast in the color-balanced stitched images. We introduce an evaluation metric to assess the extent of contrast preservation between the images before and after dodging. In our experiments, we utilize 116 Gaofen-3 and Sentinel-1 SAR images for dodging stitching and compare our proposed method to traditional dodging methods. The proposed method outperforms traditional methods both in terms of visual effects and evaluation metrics.

With the explosive growth of the number of videos on surveillance and the Internet. How to intelligently detect the large amount of video data has become a core task in the current video security surveillance system and network security, which is of great significance to urban security control and network environment governance. This paper introduce the idea of metric learning and design a video abnormal behavior detection network EBML applied to small sample size to address the problem that specific abnormal behavior cannot be detected and abnormal behavior samples are scarce in most methods. Firstly, EfficientNet is used as the backbone network for feature extraction, then, bringing in an improved BiFPN to mitigate the feature loss caused by too many networks and information loss, and enhances the ability to fuse low-level semantic information. Finally, the scalable Cosine distance metric is introduced into the Softmax of BiFPN-EfficientNet, which makes the gap between similar features shrink continuously and the gap between dissimilar features increase continuously, thus improving the accuracy of video anomalous behavior detection. By testing on the VCAD dataset, the accuracy reaches 96.3% and the F1 value is 0.94, which has an increase of 7.8% and 0.09 respectively, compared with the benchmark model EfficientNet. Experiments show that the algorithm can better balance speed and accuracy in video anomaly detection and can meet the demand of video anomalous behavior detection under small samples.

A novel rate control algorithm for immersive video depth map coding is proposed to ensure the immersive video quality in the context of limited bandwidth. The algorithm is developed based on the frame level approach for Versatile Video Coding (VVC) and is optimized using the open-source encoder implementation project, VVenC. Specifically, by considering the content characteristic of different videos, the rate-distortion (R-D) curve is refitted using hyperbolic model for the rate control parameters α and β. And the original α and β values of the depth map are modified to better guide the bit allocation. Experimental results indicate that the proposed rate control algorithm outperforms the VVenC 0.3.1.0 frame level rate control algorithm, achieving an average 4.2% BD-rate saving and 1.616dB gain under random access (RA) configuration, which is superior to existing rate control schemes for immersive video.

Blind source separation (BSS) refers to obtain source signals by means of their linear commixture with unknown mixing channel. Existing BSS methods mainly rely on the basic hypothesis that the source signal is non-Gaussian, which leads to its inability to separate the mixed Gaussian signals, and severely limits the application of BSS. In order to solve the above problem and the shortcomings that cosine similarity will encounter in the measurement of dynamic similarity of signals, this paper utilizes the generalized Jaccard index to quantify the dynamic similarity of signals, and then proposes the BSS method based on dynamic similarity of signals to realize the separation of nonlinear chaotic Gaussian signals. This method introduces the exponential function consisting of dynamic stationarity factor and independence factor of signal as the cost function of blind source separation, and then employs the imperialist competition algorithm with fast convergence speed to solve cost function to achieve the separation of signal. The effectiveness of proposed method is verified by conducting simulation experiments on the synthesized nonlinear chaotic Gaussian signals and ECG signals. Simultaneously, in comparison with the BSS method based on cosine index, the experimental results show that the proposed method has smaller cross-talking error and better separation effect.

For research at home and abroad, and to ensure the accuracy and real-time of the acquired images, to achieve the function of moving target detection, a digital image processing application based on FPGA - moving object detection is proposed, and the image acquisition and image processing two parts are studied, the image acquisition is OMINIVISION OV5640 image sensor, which comes with an embedded microprocessor to support image scaling and translation. By the programmable device FPGA to receive and send data to the FPGA external large-capacity SDRAM memory for storage, image processing is selected Intel Cyclone IV series FPGA chip, for the image acquisition process of a variety of random noise, through the verilog HDL hardware description language written RTL register transmission level code, to achieve image acquisition and storage, average filtering, expansion algorithm, corrosion algorithm, morphological operation closed calculation design The processed image is used for moving target detection and VGA display display. The system is mainly composed of camera acquisition module, SDRAM controller module, display driver module, image processing module, through design simulation, the results show that the system realizes real-time acquisition and processing of digital images, reaching 640*480 resolution display of 30 frames per second.

Current research on multi-label image classification mainly focuses on exploring the correlation between labels to improve the classification accuracy of multi-label images. However, in the existing methods, the label correlation is calculated based on the statistical information of the data. This label correlation is global and depends on the data set, and is not suitable for all samples. In the process of extracting image features, the The characteristic information of small objects is easily lost, resulting in low classification accuracy of small objects. For this reason, this paper innovatively proposes a multi-label image classification model based on multi-scale semantic attention and graph attention network. vector, followed by feature fusion to enhance the feature information of small objects, and then use the self-attention mechanism in the graph attention module to adaptively mine the correlation between categories in the image, and propose an attention regularization loss. The mAP of the model on the two public datasets of VOC 2007 and MS-COCO 2014 reached 95.5% and 83.4%, respectively, and most of the indicators are better than the existing state-of-the-art methods.

Recently, telemedicine diagnosis via the Internet has been widely used. The transmission of patients' private data over the network is subject to threats such as tampering, forgery and theft, and unauthorized theft and modification of others' data can cause great distress to patients and physicians. To protect the security of medical images. In this paper, a new encryption method for color medical images is proposed. The method uses 5th order orthogonal Latin square based encoding, DNA arithmetic and zigzag ordering to dislocate and diffuse the images. Its security is tested in terms of correlation coefficient, key sensitivity and noise attacks. The results show that the method has a high level of encryption for color images, can achieve lossless decryption, and outperforms existing methods.

Infrared images provide brightness information of targets, while visible images provide detailed information about targets. The fusion of infrared and visible images can provide the most comprehensive description of the scene. Existing deep learning-based fusion methods mostly utilize convolutional neural networks (CNNs) to extract local features for fusion, but they do not exploit the global contextual relationships of the images. On the other hand, using a complete Transformer to process the segmented image does not leverage the local features effectively. In this paper, we propose a dual-branch fusion algorithm combining CNN and Transformer. The dual-branch adopts a parallel structure, allowing for maximum utilization of both local features and global representations. By processing the sequence features generated by the Transformer branch, we ensure their resolution and dimension are consistent with the local features for fusion. The fusion strategy employs a combination of L1-norm and averaging, and our model is trained in two stages. Comparative experiments with various fusion algorithms demonstrate that the proposed algorithm achieves good fusion results and shows improvement in objective metric evaluations.

Unmanned aerial vehicle (UAV) synthetic aperture radar (SAR) is an efficient remote sensing system. It has been used widely in both the military field and civilian field. However, when it works in spotlight mode, the accumulated motion errors can be heavy over the long synthetic time, inevitably causing two-dimensional (2-D) defocus. In this paper, a novel two-step autofocus algorithm based on the coarse imaging result is developed. The estimation of azimuth as well as range error is operated in the range-compress azimuth-frequency domain and the compensation is completed with the 2-D frequency spectrum correction. Simulation experiments verified the effectiveness of the proposed method.

In order to improve the detection capability of typical non-cooperate targets, a facet-based synthetic aperture radar (SAR) imaging algorithm, and a SAR image target detection model are presented in this paper. At first, the shooting and bouncing ray (SBR) method was utilized to calculate the backscattering coefficient of each facet on the typical target surface. Then, based on the radar echo generation method and SAR imaging algorithm, the SAR images of the targets can be obtained by simulation. Therefore, a SAR image dataset can be established containing simulation results under different conditions. Finally, combined with the most recently proposed YOLOv7 deep learning model, the feature learning and training based on the target SAR dataset are realized. Compared with the previous original YOLOv5 and improved YOLOv5 networks, experimental results show that YOLOv7 performs better in precision and efficiency under the same conditions, which provides a concrete foundation for future research.

The current conventional high-voltage transmission line monitoring method mainly obtains line parameters through fiber optic sensing equipment, and analyzes the parameters to grasp the actual load situation of the line, which leads to poor monitoring effect due to the lack of design of load flow threshold. In this regard, a monitoring method for high-voltage transmission lines based on self-assembled MAC access is proposed. By constructing the reward function, the routing monitoring algorithm is designed and the image monitoring module structure is designed. Combined with the heat balance equation, the line load flow value is calculated and the judgment of the line load condition is realized by setting the flow threshold. In the experiments, the monitoring performance of the proposed method is verified. The experimental results show that when the proposed method is used for transmission line monitoring, the system data loss rate is low and has a more desirable monitoring effect.

Screen content video (SCV) distinguished from traditional natural video by its unique features, such as rich texture edges and flat areas. However, existing rate control algorithms are better fit for natural video than for SCV. A novel rate control (RC) algorithm for screen content coding (SCC) in Versatile Video Coding (VVC) is proposed in this paper. Specifically, spatial-temporal features at both CTU and frame levels of adjacent original and reconstructed frames are calculated based on 3D-LOG filter to guide bit allocation in CTU level. Considering the visual characteristics of human vision system, the rate distortion (R-D) model is redesigned. Extensive experimental results demonstrate the effectiveness of the proposed method, which improves the R-D performance and the accuracy of the bit-rate control.

Image inpainting aims to fill damaged regions with non-damaged regions and semantic reasonableness while ensuring consistency of an image. The result of inpainting often suffers from smooth edges and blurred details when faced with larger and more complicated damaged regions. In this paper, an end-to-end dual stream network that fuses the texture and structure features, aiming to restore intricate details in filled regions is proposed. For details enhancement, gated convolutions are introduced to pick valid pixels, reducing blur in damaged regions; For more comprehensive features representation, multi-scale parallel dilated convolutions are used to fuse features from different receptive fields and positions in the image. Extensive experimental results on three common datasets demonstrate the superiority of the proposed network in terms of quantitative and qualitative evaluation.

Underwater images play a crucial role in underwater exploration tasks. However, due to the unique physical and chemical properties of the underwater environment, underwater images often suffer from issues such as low contrast, color cast, and blurriness. To address these challenges, this paper proposes a dual-path fusion model (UW-DSFNet) for underwater image enhancement. The model aims to extract both color and texture features from underwater images comprehensively, utilizing spatial and frequency domains. In the spatial domain path, a low-complexity NAFNet network is employed along with gate residual and GELU activation functions to extract color features from the images. In the frequency domain path, an MLP framework is utilized, and Fourier transform is applied to obtain frequency domain texture feature maps. Finally, the features extracted from the spatial and frequency domains are fused, followed by a detail enhancement process. Experimental results demonstrate that the proposed model effectively enhances underwater images, producing clear and visually appealing results with rich colors.

With the rapid advancement of information technology, the modern battlefield is characterized by a highly complex electromagnetic environment. Radar radiation sources exhibit wide-ranging parameter variations and strong random characteristics, presenting formidable challenges to the signal selection of radar radiation sources in missile-borne countermeasure systems. This paper addresses the issue of reliable identification and selection of radar source signals by on-board countermeasures systems. Through the analysis of source signal characteristics, the Smooth Pseudo Wigner-Ville Distribution (SPWVD) method is employed for time-frequency analysis to extract the time-frequency features of the source signals. Furthermore, a lightweight network based on SqueezeNet is implemented to achieve high-precision source signal selection. The results demonstrate that, when the SNR of the source signals is greater than 0dB, the network model achieves a recognition accuracy above 94.59%. The selection accuracy is comparable to that of the Convolutional Neural Network (CNN), thereby meeting the requirements of on-board countermeasure systems for reliable selection of radar source signals. The analysis confirms that under low signal-to-noise ratio conditions, noise significantly affects the network's selection accuracy by impacting the time-frequency clarity of the modulation signals.

The building in some towns or villages have been in disrepair for a long time, and there are hidden safety hazards. However, the procedure for inspecting the risks associated with construction is somewhat antiquated, and considerable personnel and material resources are squandered. Therefore, enterprises hope to make it easier to inspect the hidden dangers of buildings by making a building safety inspection system that can be used more convenient. The building safety inspection system should help enterprises save the manpower and time cost of inspecting hidden dangers, complete the government's entrustment more efficiently, and speed up the enterprise informatization process. It also saves the cost and time of manual statistics, facilitates the process of identifying hidden dangers and management. In this paper, we propose a building safety inspection system with modified ResNet50. This system adopts FastAdmin framework technology to realize the operation process of the building safety inspection, and ECharts technology is used to display charts. Furthermore, the modified ResNet50 is a 50-layer ResNet by using multiple layers, such as convolution layers, pooling layers, and fully connected layers to automatically inspect the building quality. From results, the average testing accuracy for ResNet50 is 99.5%, and the accuracy is better than convolution neural network (CNN). The proposed system can meet the enterprise requirements.

For the case of imperfect channel state information (CSI) in the downlink multi-carrier non-orthogonal multiple access (MC-NOMA) scenario, we propose a dynamic resource allocation method, which is committed to maximizing the sum rate of users in the NOMA scenario and maximizing the efficiency of the NOMA system performance. We decompose the combinatorial optimization problem, and first derive the closed-form of power allocation under one single subchannel. According to this result, we use a deep reinforcement learning (DRL) method to solve the exhaustive problem of user pairing, and propose a user pairing algorithm. Simulation results show that our method is more efficient and accurate than traditional methods.

Blockchain, as a distributed database, has been wildly used in real-world applications. Different systems based on blockchain form data silos. Therefore, cross-chain data transmission is required to realize value transfer in different systems. In the process of cross-chain data transmission, authentication is important for data security. In this work, we design an auditable anonymous authentication scheme for cross-chain data transmission. In our scheme, the sender can generate anonymous authentication message on the data. Then the receiver of the target blockchain knows the data is sent by a legal user of the original blockchain, but cannot know the real identity of the sender. By using Trusted Execution Environment (TEE) and pseudo identity methods, our scheme is more efficient than the similar anonymous authentication schemes in blockchain.

Unmanned Aerial Vehicle (UAV) into Mobile Edge Computing (MEC) systems can effectively expand their flexibility and coverage. The current research trend in UAV-assisted MEC is mainly focused on optimizing flight control and minimizing energy consumption. However, there is a lack of research on task scheduling in UAV-assisted MEC scenarios. This paper proposes a two-stage task scheduling method that minimizes the task processing cost while optimizing task execution order to reduce the execution time of tasks. The results of the study show that the proposed algorithm outperforms other baseline methods in terms of completing all tasks with the shortest execution time. This further validates the efficiency of the proposed algorithm and its potential for improving the system efficiency of UAV-assisted MEC.

The study of signals defined on irregular domains has garnered significant attention, and the field of graph signal processing has emerged as a way to model such signals using underlying graph structures. Vertex frequency analysis is a crucial tool for graph signal analysis and representation. Finding different kinds of frames via the windowed graph fractional Fourier transform (WGFRFT) is necessary to extract the characteristics of graph signals. To overcome this issue, we introduce the concept of the dual of WGFRFT frame and use it to develop a new reconstruction formula for graph signals. Additionally, we propose fractional shift frame. Finally, to highlight the utility of our proposed frames, we present some examples which use our fractional shift frame to extract the spectrum of graph signal.

This paper proposes a fault diagnosis model based on the combination of continuous wavelet transform and improved first-layer wide convolutional neural network (CWT-IWDCNN). The model first performs continuous wavelet transform on the data containing a lot of noise to extract its characteristic data, and converts the original data into a time-frequency map. Then, the time-frequency diagram is fed into the IWDCNN model to obtain the diagnosis result. The proposed CWT-IWDCNN has the following advantages. 1) The continuous wavelet transform can effectively extract fault features; 2) The first layer of the convolutional neural network uses a wide convolution kernel to suppress high-frequency noise, and other layers use small the convolution kernel improves the domain adaptive ability of the model through nonlinear mapping to improve the accuracy of diagnosis; Finally, in order to test the performance of this model, the CWRU dataset is used for experimental verification. The experiments show that the proposed fault diagnosis method is better than WDCNN in most cases in terms of accuracy and robustness.

Rapid extraction of ship target information in Synthetic Aperture Radar (SAR) images plays an important role in sea surface monitoring and military prevention. However, the existing detection algorithms have disadvantages such as large model volume and slow detection speed, which are not suitable for the requirements of future star-earth integrated target detection. To solve these problems, this article proposes a SAR ship target detection method based on the improved Nanodet algorithm. To solve the problem of multi-level feature map fusion, the Ghost-pan module is added to the network to enlarge the receptive field and better fuse multi-scale features. At the same time, Resnet18 is used instead of the original backbone network, and depth-wise separable convolution is used instead of ordinary convolution to reduce the model parameter volume and improve detection efficiency. Conducted ablation experiments on the SAR dataset, and the results show that the proposed method achieves better accuracy and faster detection speed.

In order to monitor the video state changes of safe human settlements in real time, a Mean Shift algorithm is proposed. The video monitoring images collected in real time were preprocessed with enhancement and denoising, and the gray distribution was extracted to form the feature matrix of the image. The monitoring images collected in real time by the monitoring system and the monitoring images in normal state were classified by Mean Shift algorithm. Experiments show that Mean Shift algorithm has fast convergence speed and high accuracy, and can effectively screen out abnormal images and improve retrieval efficiency.

The nonlinear action of acoustic waves in a medium generates acoustic radiation forces, which can be used to manipulate particles and droplets in gases. The acoustic manipulation technique can manipulate larger objects in media without contact, and therefore has been widely used in chemical analysis, contactless transport and bioreactors. In this paper, a subwavelength straight tube acoustic manipulation device is designed to generate a standing wave acoustic field inside a finite length circular waveguide by exciting a single low-order mode with two small transducers. It is found that the position of the particles inside the subwavelength tube can be controlled by switching the frequency, so that the particle transport inside the subwavelength tube based on the frequency control of the acoustic field can be realized. This work then uses a flexible PVC pipe and successfully controls the particle movement inside the PVC pipe with different bending angles, verifying the applicability of the method of subwavelength pipe particle transport based on acoustic field frequency control. Finally, this work further designs an acoustic particle collision device based on the above method using two sets of acoustic sources. This work further expands the use scenario of subwavelength pipe-enhanced acoustic tweezers, which is expected to deepen the physical understanding of the acoustic field on matter interactions and develop new miniaturized acoustic manipulation devices inside pipeline structures.

Scattering centers are important features of targets at high frequency regions and the geometric theory of diffraction (GTD) scattering center model is the typical one to describe scattering centers. Therefore, it is quite vital to estimate parameters of GTD model accurately. The classical multiple signal classification (MUSIC) algorithm can effectively estimate these parameters at high signal-to-noise ratios (SNR) but it suffers from poor parameter estimation performance at low SNR scenarios. To solve this problem, we propose a modified MUSIC algorithm which enhances the noise robustness. The modified MUSIC algorithm construct a new total covariance matrix R by averaging the auto-correlation matrix of the original back-scattered data and the auto-correlation matrix of its conjugate data. Then, we take even powers of R ,which can broaden the differences between the eigenvalues of noises and signals and avoid overlapping spectral peaks. The theoretical computational complexities of the main modified step are discussed in this paper. Simulation results verify that the proposed algorithms achieve superior accuracy in parameter estimation of the GTD model and obtain better noise robustness. What is more, a target recognition method based on the GTD model and the artificial intelligence algorithm are proposed in this paper. Simulation results validate the effectiveness of this method.

Vehicle pedestrian detection is a key aspect in driver assistance systems, which need to accurately detect all vehicle pedestrian targets on the roadway in order to ensure driving safety. To solve the problem of low accuracy in vehicle pedestrian target detection, this paper proposes a vehicle pedestrian detection method based on the improved YOLOv5 algorithm. In this paper, the initial anchor boxes of the dataset are re-clustered by the K-means clustering algorithm, and the CIOU loss function and DIOU_nms, are applied to the YOLOv5 algorithm to improve the target recognition effect and reduce the false and missed detection rate of small targets. The experimental results show that the mAP@0.5 of the improved YOLOv5 algorithm is improved by 1.85%.

At present, most of the semantic segmentation algorithms for road scenes are aimed at visible images, but the visible images collected in dark and other special conditions will seriously affect the accuracy of semantic segmentation, so to enhance the ability to understand the night road scene, this paper obtains the infrared image of the night road scene, based on the improved DeepLabV3+ network structure, so that the network takes into account both segmentation accuracy and real-time performance. First of all, MobileNetV2 is used as the backbone network to reduce the amount of computation and the number of parameters in the network. Secondly, the channel and spatial attention modules are used in the encoding and decoding stages to obtain rich context information, filter background information, and reduce the loss of detailed information. Finally, the tightly connected atrous spatial pyramid pooling (ASPP) module is introduced, which not only expands the receptive domain but also extracts more abundant multi-scale features. The experimental data show that the average intersection-merge ratio of this algorithm is obviously better than that of the DeepLabV3+ algorithm on infrared data sets, and the average intersection-merge ratio can reach 81.45, showing better accuracy and real-time performance.

Recently, metarsurfaces working in terahertz wavelength region has attracted considerable attention from researchers, which could provide an effective way of advancing the performance of the terahertz bio-sensors. In this paper, we propose a metasurface array which consists of a square ring with four gaps at each corner and a wire inserting parallel to x axis. Simulation results show that one additional resonance is excited at 2.173 THz due to the introduced asymmetry. According to the calculation, the sensitivity of the additional resonance can reach 606 GHz/RIU. Meanwhile, the Q factor and FOM value can also reach 50.53 and 14.09 respectively. Due to the excellent performance, the proposed metasurface array can be effectively used for terahertz sensing application.

With the development of technology, the intrapulse modulation technology of radar signal is becoming more and more complex. In this paper, a method based on multilayer convolutional neural network (MCNN) is proposed to identify the intrapulse modulation mode. Firstly, five signal modulation modes (conventional pulse signal, linear frequency modulation signal, sinusoidal frequency modulation signal, phase shift keying signal and frequency shift keying signal) are established. By changing the values of carrier frequency, pulse width and repetition frequency, five kinds of training data are generated with a certain signal-to-noise ratio (SNR). Then, five trained MCNNs are obtained by training five kinds of training data with a MCNN. Further, the recognition performance of different trained MCNNs is studied with the test data generated under different SNRs. Finally, the simulation shows that the mode with parameter variations has the best recognition performance.

Generally, there always are some surface defects, just like scratch, imprint, surface inclusion etc, in the surface of the strip steel. So it is difficult to determine the edge in the process of strip steel width measurement. Facing this problem, this paper proposes a method of edge detection in the strip steel width measurement. Firstly, the gradient calculation and the non-maximum suppression method are used, and then the statistical decision method is designed to detect the exact position of the strip steel edge. This method has achieved good results in the strip steel edge detection.

With the growth of the Internet and the age of information explosion, more and more people need network drives to store and share the resources they find online. While downloading files to a local hard drive can ensure data security, this also brings with it many inconveniences, such as the need to share with others using third-party websites or communication software, or the need to transfer files using a removable hard drive. As a result, personal network drives have emerged on the market to make data sharing easier by transferring files to a server. This project will retain the basic functions of a traditional network drive system, such as file uploading, downloading, and deletion, while taking into account the security of the data and using deep learning-based DBN algorithms for image recognition, using artificial intelligence and image recognition techniques to detect pirated and illegal image material. Through the combined application of the above technologies, an efficient personal web-disk system is designed and implemented.

Interleaved antenna arrays are widely used in radar tracking, navigation, wireless communications, and location, etc. In order to make use of a given antenna aperture of concentric rings array (CRA) efficiently, an interleaved method for interlacing uniform CRA is proposed in this paper based on Subarray element Excitation-Matching (SEM). Due to Fourier-Bessel transform mapping between array factor and uniform CRA antenna excitation, in our method, the array factor of CRA is first calculated and its sidelobe region is then amended accordingly to meet user-defined sidelobe level threshold. In the next steps, the updated antenna element excitation coefficients are acquired with 2-D fast Fourier transform (FFT) and cubic interpolation method. Finally, with a careful exploration of the excitation distributions, the antenna positions of thinned subarrays are deliberately chosen in an alternating manner, which ensures that a similar pattern can be achieved with different thinned subarrays. The performance of the proposed interleaved CRA design method is evaluated by numerical simulation.

Detecting the key points of the peri-implant alveolar bone is the most crucial step in automatic clinical diagnosis. To achieve keypoint detection accurately and efficiently, we propose an end-to-end convolutional neural network-based method for detecting key points of the peri-implant alveolar bone. The method uses a High-Resolution Network (HRNet) for feature extraction and regards type and identity labels as different attributes and achieves single-stage pose estimation via Attribute-Disentangled Heatmap (ADH) prediction. ADH is end-to-end trainable and presents better robustness to accumulated errors in two-stage models. It jointly optimizes keypoint type and identity prediction in a unified model, hence enjoying better efficiency and compactness. We compared the proposed method with the state-of-the-art keypoint methods ResNet-50, Mask-RCNN, and Hourglass, and the experimental results show the superiority of our method. The mean Average Precision (mAP) value achieves 85.6% in mAP, which outperforms all the compared methods.

Synthetic aperture radar (SAR) has a special ability to work in any type of inclement weather, and is a very suitable tool for Ocean surveillance. Scene classification is an essential pre-task of other computer vision tasks for ocean monitoring. It is of great importance to develop scene classification technology of SAR sea images. Due to the excellent feature representation abilities of neural networks, the deep learning-based methods are far superior to the traditional methods based on manual features in scene classification task performance. Many lightweight classification networks have been proposed to improve the inference speed of the networks. But in comparison with ordinary CNNs, the lightweight networks have slightly lower accuracy for scene classification tasks. So in this article, we proposed an improved lightweight Convolutional Neural Network for scene classification of SAR sea images. First, in order to meet the real-time performance, we choose MobileNetv1 as the original classification network in this paper. Then, to compensate for the lack of accuracy, we use 1D asymmetric convolution kernels to strengthen each layer of the depthwise convolutions in the network. Finally, after training time, we merge the linear calculations of each layer of the network to convert it into the original structure. The experimental results show that the modified model has obtained an accuracy improvement than the original one on the scene classification of sea SAR images without extra computation.

This paper presents a velocity estimation method to addresses the steering vector mismatch of angle measurement and adaptive coherence estimation (ACE) in mDT space-time adaptive processing (STAP). The proposed velocity estimation method is dual to the adaptive monopulse in space-domain, and it can be conducted in a simple process framework. Numeric results show that the angle estimation errors are well reduced and detection aided with ACE are improved.

In this paper, a filtering antenna array based on substrate integrated waveguide technology is designed. The antenna works in the Sub 6G band of 5G, and the array effect is formed by etching slots on the upper surface of the cavity at specific locations to excite the patch, while the filtering effect is generated because of the SIW cavity and patch coupling, forming a high gain filtering antenna. Simulation results show that the filter antenna has 3% bandwidth, peak gain greater than 10dBi in the operating band, and front-to-back ratio greater than 20dB.

To address the problem that the geometric connection of the channels was ignored and the segmentation accuracy was not enough in existing point cloud part segmentation, a point cloud part segmentation model based on channel attention mechanism was proposed. First, adaptive furthest point sampling was used to determine the sampling points so that the determined local area is better representative, so that the local area could be represented better. In order to suppress the interference of useless information and capture more useful local feature details, three kinds of channel compression operators were used to perform attention pooling in the local area, so as to calculate the attention distribution of channel features. At the sametime, feature interpolation was used to classify each point to achieve the purpose of overall point cloud segmentation. Finally, the ability of the experimental model to capture channel geometric features was verified by comparing the segmentation results of two point cloud datasets with different channel dimensions. The results show that the proposed model can effectively improve the segmentation accuracy of 3D point cloud classes and parts. It can maintain the robustness of the segmentation effect in the case of different number of channels.

The "Closing and Centering" of the miter gate of the large shiplock is one of the core control process requirements of the ship-lock operation control system. It is an important guarantee for the safe and efficient operation of the shiplock. Taking the miter gate of the Three Gorges Shiplock as the object, this paper studies the principle and method of image recognition, and adopts the methods of real-time image preprocessing, filtering and denoising, threshold segmentation, algorithm detection, Hough transform straight line fitting, system calibration and mathematical conversion to realize the real-time monitoring scheme of miter gate centering door gap. The research results are helpful to realize the real-time monitoring of the closing process of miter gate of Three Gorges Shiplock, and effectively avoid the collision and extrusion accident caused by the too large closing gap. It provides the theoretical support for the navigation benefit of the Changjiang River Junction and promoting the automatic and intelligent operation level of the Three Gorges Shiplock.

Target detection techniques based on computer vision can be used for vegetable recognition and classification, which can effectively avoid the time-consuming and labor-intensive problems faced by manual operations at the large supermarket checkout lanes and produce market. In order to accurately recognize vegetable image, an improved Faster R-CNN recognition technique based on the Swin Transformer is proposed in this paper. The backbone network is replaced by the Swin Transformer to improve the accuracy and efficiency of feature extraction. The ROI Align is employed to protect the integrity of image data. Furthermore, the more effective GIoU loss function is used to simplify the training process in order to reduce the computational resources and time consuming during training. Finally, the experimental results show that the proposed algorithm has an accuracy improvement of 6.1% compared with the previous methods.

Generative adversarial networks (GANs) have achieved remarkable success in image translation tasks. However, existing models and theories are still limited in their ability to generate new fonts that combine the styles of both font domains. Most current models only perform style conversion, which often results in generated fonts that are biased towards the target font domain, such as Chinese characters. Moreover, the long training time of these models makes them impractical for real-world applications. To address these issues, we propose a novel approach for font generation based on the pix2pix and AC-GAN frameworks. Our model can learn the mapping from the source font domain X to the target font domain Y without biasing the generated fonts towards either font domain. During training, the generator is optimized to produce high-quality new fonts that combine the styles of both font domains, while the discriminator is trained to distinguish between real and generated fonts. Experimental results show that our approach can generate 2555 new fonts with both font styles from a training set of 943 characters in a short time. The generated fonts are visually appealing and combine the styles of both font domains, achieving the goal of fusing styles and generating new fonts.

Wearable flexible sensors are being increasingly designed for monitoring human health states. In this study, we optimized the preparation process of paper-based graphene and designed amplifier filter circuit to sense weak pulse signals with high sensitivity (0~300Pa) and fast response time (<1ms). We further used a convolutional neural network (CNN) as the recognition method to analyze the collected pulse signals. We successfully achieved the recognition and classification of three signals (Cun, guan, chi, and Sport) with an accuracy of 85.79% in the training set and an accuracy of 80% in the testing set. Our research offers new possibilities for the wide application of paper-based graphene sensors in medical diagnosis and health monitoring.

Ink painting is a typical representative of Chinese painting. With the acceleration of globalization and the continuous development of deep learning and other technologies, Chinese ink painting urgently needs to develop into digital development. Generated adversarial network has been widely used in the field of image style transfer, among which, ChipGAN model is specifically aimed at the study of Chinese ink painting style transfer. In previous studies, we have tried to stylish the animated film ink painting, but there is no transfer learning under the ChipGAN model, and the comprehensive transfer effect evaluation is lacking. This paper will be for the animated film "big fish haitang", select five different ink style style data set, using ChipGAN model for ink style transfer, and use the SSIM image quality evaluation index and HSV color difference value from the content and style of transfer image for a more comprehensive evaluation, to explore the suitable for Chinese ink painting image transfer evaluation method lay the foundation.

In the pharmaceutical industry, the label on vials of antibiotics contains crucial information about the medication. The quality of the label's information determines significant issues such as drug classification, storage, safe usage, and also affects the economic benefits of pharmaceutical manufacturers. In response to the shape distortion that may occur in cylindrical labels, this paper proposes a cylindrical back projection correction algorithm. The adjacent label images are stitched together after cylindrical back projection to create a complete cylindrical label image. The corrected and stitched label image is then used as the input dataset for defect detection using the ResNet-18 residual network. A comparison is made between the defect detection results obtained from the corrected and uncorrected label images. The feasibility of the algorithm is validated by comparing the accuracy of the two sets of experimental results. The experimental results demonstrate that the proposed algorithm improves the accuracy of cylindrical label defect detection.

This paper builds an end-to-end custom object detection model that could translate Chinese numbers’ sign language in real time based on deep leaning. The main work of this paper is as follows, collect images for deep learning using OpenCV, label images for sign language detection using LabelImg, setup Tensorflow Object Detection pipeline configuration, and use transfer learning to train a deep learning model. At last, detect sign language in real time using OpenCV.

The recognition of road targets represents one of the central technological advancements within intelligent transportation systems, aimed at addressing the critical problem of urban congestion. In light of this, the present paper introduces a road target recognition algorithm, MOLO, specifically designed to navigate complex traffic environments, and subsequently deployed to facilitate detection via mobile devices. MOLO comprises two distinct components: a feature extraction network and a feature fusion network. The former of which, builds upon the structure of SSD network, integrates the lightweight receptive field expansion module RFB-L, whilst the latter introduces a pixel and channel information fusion module termed Fusion, thereby enabling the fusion of both shallow and deep features, ultimately enhancing the overall detection efficacy of the model. The model transformation process uses quantization and inverse quantization strategies between fixed-point and floating-point numbers, significantly reducing the number of parameters, all whilst preserving the accuracy of the detection performance on mobile devices. Experiments show that MOLO, when deployed to mobile, achieves an inference speed of 23.66ms and with 98.96% detection accuracy on the homemade road dataset Hohhot_city, and also shows superior performance in actual road detection.

The advancement of computer technology has facilitated the provision of a wide range of services to customers through computer networks. However, the security of these networks is constantly threatened by malicious actors who employ various means to launch attacks. Denial-of-Service (DoS) attacks are a common type of network attack that poses a significant risk to many computer information systems. In this research paper, we aim to investigate how the efficiency of DoS attackers is affected when they have limited computer resources. The findings of our experiments confirm that DoS attackers face challenges in executing efficient attacks when their computers lack sufficient CPU resources. By understanding this limitation, it becomes possible for the targeted entities to choose appropriate defense strategies when faced with such attacks.

Aiming at the problem of Direction of Arrival (DOA) estimation of coherent mixed targets in the far field of uniform linear array, the deep convolutional neural network and deep convolutional self-encoder are designed by combining deep learning with DOA estimation of coherent sources. The deep convolutional encoder is trained by comparing the difference between the received covariance matrix of the independent source array and the received covariance matrix of the coherent source array under the same condition, so as to realize the process of decorrelation, and then DOA estimation is carried out. The simulation results show that both methods can extract spatial features sufficiently, improve the accuracy of DOA estimation and reduce the complexity of the algorithm, and the method based on deep convolutional self-encoder has better performance.

In recent years, Transformers have demonstrated significant potential in segmentation tasks. Self-attention are generally believed to be crucial for improving the performance. However, these excellent results are based on large-scale imaging datasets, and there is not suitable for low data regimes such as lesion segmentation task. In this paper, we proposes adopting a few-shot data-friendly pre-trained convolution kernel to replace self-attention as the feature mixer. Specifically, we have designed a new information transmission layer which extracts feature information through convolution with initialization information parameters, and performs feature mapping through MLP. To verify its effectiveness, we conducted experiments on two low data sets, CVC-ClinicDB and 2018 Data Science Bowl challenge dataset. Surprisingly, we proposed scheme achieves a Dice Similarity Coefficient (DSC) of 93.17%, which is 19.18% and 12.64% higher than the two baselines TransU-Net and SegFormer with self-attention, respectively. Therefore, PCFormer is encouraging and can be relied upon for few-shot lesion segmentation.

This paper presents a multi-angle imaging system that integrates Multiple-Input Multiple-Output Synthetic Aperture Radar (MIMO-SAR) with the Multi-Angle Synthetic Aperture Radar (MA-SAR) in the context of spaceborne SAR's spotlight operation mode. The system addresses the issue of narrow swath coverage in traditional multi-angle imaging. It significantly relaxes the requirements on the pulse repetition frequency (PRF) compared to the Doppler bandwidth of the echo signal and achieves wide-swath imaging by resolving azimuth ambiguity. Simulation experiments demonstrate the successful integration of MIMO-SAR with multi-angle imaging, showcasing excellent imaging performance.

In wireless laser communication systems, beam acquiring, pointing, and tracking (APT) technology is the key to establishing and maintaining a laser communication link. In order to solve the problem that the capture cameras used in APT systems usually have a small field of view and are inefficient in scanning the uncertain area and capturing the target spot, which leads to the inability to precisely locate the spot position, an image processing-based omnidirectional laser communication spot image detection technology solution is designed under the condition of using a large field of view of the capture camera, which has the characteristics of strong anti-interference capability and high detection accuracy. Firstly, the block design was simulated on matlab platform, and then the APT system of omnidirectional laser communication was built based on the panoramic camera with large field of view and two-dimensional turntable to verify the scheme. The scheme is implemented in 0.1° steps to sample the laser spot within 180° of the horizontal and tilt directions and obtain the spot coordinates. The spot position information is sent to the two-dimensional turntable through the serial port, and the 0.1° pointing of the turntable to the laser light source is accomplished, which verifies the feasibility and accuracy of the system solution.

The latest video coding standard, the Universal Video Coding Standard (VVC), uses new coding tools to greatly improve compression efficiency. However, the adaptive QP module in the coding framework ignores the characteristics of the human visual system HVS, resulting in coding results that differ from the real perception of the human eye. To incorporate visual perception into QP selection, this paper proposes an adaptive QP offset algorithm based on visual perception metrics. The first step is to design a QP offset algorithm based on the luminance, frequency, and time domain characteristics of the HVS. The visual perception index is designed to reflect the real perception of human eyes, and then the index is used to guide the CTU into different levels according to visual sensitivity. Experiments show that the algorithm achieves 0.74% performance improvement in VMAF - based BD-Rate and about 8% improvement in VMAF score compared to the original VVC.

Automatic traffic signal detection has great significance to the development of autonomous driving technology and vehicle warning of dangerous driving behavior. Detecting traffic lights by analyzing images obtained from mobile device cameras is a feasible solution. In this paper, a new traffic signal detection system with high robustness and real-time requirements. Firstly, the HSV color segmentation method is used to segment the original map. Then, the aspect ratio information and area information of traffic lights are utilized for secondary filtering. The morphological processing method is used to filter the background in the image. Finally, intercept the crucial area to obtain the final significance segmentation image of preprocessing. Combining the salient feature preprocessing based on traditional image processing methods with the target detection model based on deep learning, the algorithm of traffic signal detection and recognition is improved.

Space-Ground Integrated Network (SGIN) is a multi-domain integrated network with a large coverage area, which can better meet the ubiquitous communication needs in the network and provide users with better services. It has become the main development direction of 6G network. The different networks integrated in SGIN are heterogeneous, which is specifically reflected in the fact that the periodic movement of satellite networks brings time-varying nature to resources, and the expansion of network scale also brings spatial attributes to resources. It can be concluded that resources in SGIN have multi-dimensional. When performing virtual network embedding(VNE) in SGIN, the existing embedding algorithms often do not consider the multi-dimensionality of resources, which may lead to problems such as spatially dispersed embedding results and inability to adapt to network dynamic changes. To address this problem, this paper uses the space-time resource tree(S-TRT) model to represent the multi-dimensional resources in SGIN, reflecting the performance of resources in four dimensions: time, space, type, and quantity. On the basis of this model, combined with multiple dimensions of resources and the spatial distribution of embedded nodes, the ranking vector of virtual nodes is established, and the dimensionality reduction sorting of each virtual node is carried out by multi-dimensional scaling method. Afterwards, the node embedding is completed by dynamically sorting the physical nodes to improve the spatial concentration of the embedding results and better adapt to the dynamic changes of the underlying network. Finally, we conducted a simulation experiment on the algorithm, and the results show that the algorithm has good performance in the request acceptance rate and long-term revenue-to-cost ratio.

Characterized as large bandwidth, low latency and high data rate, millimeter-wave apply to mobile communications is widely applied and studied. Meanwhile, the new challenge for telecommunications operators is required for millimeter-wave planning because of high propagation loss. On the basis of introducing the characteristics of millimeter wave frequency band, this paper briefly describes and analyzes two kinds of millimeter wave coverage enhancement technologies, namely reconfigurable intelligent surface and relay communication, which are research hotspots at present, in view of the limited disadvantages of high propagation loss and coverage restriction faced by high-frequency network planning. Further, this article combines the application prospect of enhancement technology and potential deployment environment information, The feasibility reference suggestions are put forward for operators to plan millimeter wave scenarios in the future.

Aiming at the problem that AKAZE algorithm has slow feature extraction speed and low accuracy in the feature matching process, this paper proposes to improve AKAZE's feature matching algorithm based on grid statistical motion. Firstly, in the feature extraction stage, the proposed algorithm uses the oFAST algorithm instead of constructing scale space to extract feature points. Then, the M-LDB descriptor is used for feature point description. Finally, the BF algorithm is used to perform coarse matching of features, and the grid motion statistics (GMS) algorithm is added to achieve the purification of matching point pairs and complete the matching. The performance of the proposed algorithm was compared with the AKAZE and ORB algorithms in the experimental fields and grayscale graph groups. The results show that the improved algorithm not only improves the matching speed, which is more than 2 times faster than the AKAZE algorithm, but also maintains a high matching accuracy, which is similar to the AKAZE algorithm.

Based on the definition of ellipse: the sum of the distances from the plane to the fixed point F1 and F2 is equal to the trajectory of the moving point P with a constant ( greater than |F1F2| ). It is extended to the study of the trajectory of the point with the distance from the moving point to the three lines and the fixed value, and the corresponding coefficients are added before the distance from the point to the three lines, so that the sum is fixed, that is k₁ ⋅ d(P,l₁)+k₂ ⋅ d(P,l₂)+k₃ ⋅ d(P,l₃) = M(M ≥ 0). With the help of drawing software to make the trajectory, based on analytic geometry and Auto CAD, a class of quadrilateral pyramid curve trajectory is obtained, which is a geometric projection similar to the conical curve. The trajectory can be intercepted on the quadrilateral pyramid with a rectangular bottom surface, and then the relationship between the variables in the curve trajectory equation is studied. In the variation, the third straight line is changed into a fixed point, and the obtained trajectory is a conic curve with the point as the focus. In addition, in the study of eccentricity, a triangle model is established to determine what kind of curve the trajectory will be. This kind of curve can be applied to the field of UAV communication. Combined with the relevant knowledge of mathematical analysis, the practicability of this kind of curve for signal enhancement and trajectory optimization is further studied from the perspectives of uniform distribution and Gaussian distribution. Finally, Monte-Carlo simulation and covariance verification are used to prove the help and role of pyramid curve in industry and information field.

In this paper, a new optimization design method for wide-area shaped multi-beam antenna is proposed. The performance of the antenna is improved by the shape optimization of the reflective surface, and the high gain and high C/I performance of the wide-area covered shaped multi-beam antenna is achieved. In the optimization, the reflector expansion coefficient is taken as the variable, and the objective function is constructed by the antenna gain and C/I. Then the differential evolution algorithm is used for optimization to obtain the best performance. The simulation results show the effectiveness of the algorithm, and the optimization algorithm also takes the engineering realizability of the product into full consideration. The measured results show that the product has better performance than the design requirements, the antenna has been launched with the satellite, and the in-orbit test and application are good.

In recent years, blockchain has developed rapidly and achieved great success in a variety of industries. However, the current shortcomings of blockchain in terms of low throughput and slow transaction confirmation have led to the difficulty of implementing and using blockchain in some highly concurrent scenarios. In this regard, blockchain technology based on DAG was born. DAG is characterized by high concurrency and high throughput due to its unique data structure, and its combination with blockchain can break the blockchain throughput bottleneck. In this regard, a DAG-based blockchain consensus algorithm is proposed to fully order all blocks, provide the total order and effectively improve the throughput of the blockchain using the division of subgraphs. The new proposed consensus algorithm is experimentally demonstrated to have significant improvement in data throughput, and it is found that SgDAG has 79 times higher throughput than Bitcoin, 36.9 times higher than Ethereum, and 5.5 times higher than IOTA, with high throughput and can use smart contracts

According to the requirements of satellite-terrestrial network communication for large manned spacecraft, an integrated satellite-terrestrial network structure suitable for manned spacecraft was proposed, and the network protocol was designed, which realized the expansibility of multi-cabin manned spacecraft network communication and the support of ground protocol. At the same time, the data security requirements under this network architecture are analyzed, and the methods of external attack protection and increasing the internal security of the network are designed. The virtual channel is used to allocate a single virtual channel to the network data link, which improves the security and reliability of the network. Finally, the correctness and validity of the design are verified by the flight of the Tianhe core module in orbit. The design can provide reference for the future network communication design of manned lunar landing and deep space exploration.

Mine electric locomotive driverless system is an important part of transportation operation in the national mine intelligent construction, which makes the mine safe production reduce staff and increase efficiency. This topic is based on the relevant guidance documents of the State to accelerate the intelligent development of coal mines and the Ministry of Industry and Information Technology on promoting 5G to accelerate the development. Through the investigation of the current situation of the network environment of the unmanned driving system in mines and mines at home and abroad, and with the implementation background of the remote control system of the underground electric locomotive of Anhui Development Mining Co. Using the narrow and long terrain of the underground roadway, 5G and WIFI network distribution characteristics, Reasonably plan and configure the network information of relevant nodes of the system, and propose a communication link control strategy method of mine driverless electric locomotive based on 5G and WIFI hybrid network environment. The trial operation and opening of the system also demonstrated the feasibility of the control strategy.

D2D (device to device, D2D) communication is one of the key technologies in 5G wireless networks, which is very effective in improving the network resource utilization and reducing the latency of proximity communication. However, D2D short-range communication technology can bring interference to cellular systems while improving spectrum resource utilization. In this paper, we propose a joint power control and channel selection communication resource allocation scheme with the aim of maximizing system throughput while reducing inter-user interference. First, all optimal transmit power finite sets are derived using convex function theory under the constraint of ensuring the QoS (quality-of-service, QoS) of system users; then, the classical KM (Kuhn-Munkres, KM) algorithm is used to match the best multiplexed subcarriers for D2D users to maximize the system throughput after multiplexing. The simulation results show that the proposed algorithm can effectively improve the system performance.

In this paper, a blind watermarking algorithm based on redistributed invariant integer wavelet transform (RI-IWT) and BP network is proposed. The host image is processed by RI-DWT and QR decomposition, and the watermark is embedded in the low frequency and high frequency sub-band of the host image, so as to increase the embedding capacity. Moreover, we realize the blind extraction of watermark through BP network. In the extraction stage, the host image information and part of the watermark information are no longer required, thus improving the security of the algorithm. Finally, we simulate the algorithm and compare it with the relative algorithms. Simulation results show that the proposed algorithm has high invisibility and good robustness against JPEG attacks, rotation attacks, noise attacks.

The paper give our study of a hyperchaotic synchronization erbium-doped fiber laser secure communication system with dual-channel secure transmission performance, its encoding and decoding. First, we present a signal transmission synchronization scheme, where a signal from one of a hyperchaotic erbium-doped fiber three-ring laser as an emitter transmits to an erbium-doped fiber dual-ring laser as a receiver while the emitter can synchronize with the receiver by anti-phase control technique. Second, we give an encoded signal secure transmission scheme, where two signals from two ring of a hyperchaotic erbium-doped fiber three-ring laser mask two information signals, respectively, transmit to two terminals of the receiver with an erbium-doped fiber dual-ring laser to perform two encoded signal secure transmissions in two channels. Third, we give a decoding scheme, where two carrier signals from two channels subtract two signals from two ring of an erbium-doped fiber dual-ring laser to result in decoding realizing. Then, such hyperchaotic encoding secure communication laser system can perform on dual-channel secure encoding transmission and enhance the security of classified information signals. Our numerical result indicates that the hyperchaotic secure communication laser system can perform on dual-channel secure encoding transmission, and succeed decoding in dual-channel. The hyperchaotic secure communication laser system has an important reference for people’s study of optics chaos secure communication, multi-channel secure transmission, encoding and decoding.

With the rapid development of electric Internet of Things, electric power enterprises have increasing demands on information acquisition and processing ability, power grid management ability and information interaction ability. In order to further fit the future trend of power grid construction and solve many problems such as the accurate identification of the relationship between households and changes in the low-voltage distribution network, current fingerprint communication technology has been gradually widely used. At present, it is mainly applied in the equipment such as external circuit breakers, intelligent measuring switches and smart meters. However, there is no mature testing method and testing equipment to test the performance of current fingerprint communication. Therefore, this paper mainly studies the evaluation technology of current fingerprint communication. Through the design of system evaluation architecture, the acceptance capability, emission capability, consistency and application practice of current fingerprint communication are comprehensively evaluated and tested, and the application indicators and detection indicators of various aspects of current fingerprint communication technology are studied.

Blockchain is a decentralized and tamper-proof distributed ledger technology. The consensus algorithm is one of the key technologies underlying blockchain. The Proof of Stake (PoS) consensus mechanism determines the next block producer on the blockchain based on the user's staked equity. Compared to Proof of Work (PoW), the Proof of Stake consensus mechanism solves the problem of wasted resources. However, the PoS consensus mechanism faces serious issues such as the coin age accumulation attack and the zero-cost benefit problem. Therefore, this paper proposes a Trust-Based Proof of Stake (TPoS) mechanism based on dynamic trustworthiness. TPoS divides nodes in the network into miner nodes and basic equity representative (shareholder) nodes, and assigns corresponding trustworthiness to nodes based on their participation in creating blocks. Shareholder nodes sign blocks and assign them trustworthiness, and finally compete for the weight of trustworthiness obtained by blocks to go on the chain. In addition, this paper analyzes the attack cost and system response to bribery attacks and common equity accumulation attacks. The results of simulation experiments show that the TPoS mechanism has significant advantages over traditional Proof of Stake mechanisms in dealing with bribery attacks and equity accumulation attacks.

Currently, with the continuous in-depth research and application of blockchain access control, security issues on the blockchain have become a focus of attention. Based on CPABE, this paper proposes a trusted and secure blockchain access control scheme based on ciphertext policy. Firstly, a decentralized attribute-based encryption algorithm (DABE) is adopted to achieve distributed calculation of user attribute private keys, effectively solving the problems of high trust cost and single point of failure caused by the key center generating private keys in traditional CPABE. At the same time, a private key consensus verification protocol based on zero-knowledge proof is designed to ensure the correctness and security of user attribute private keys without leaking private key information. Through the analysis of on-chain security and experimental simulation, the results show that this scheme has better performance while maintaining high security and is more suitable for distributed access control with large attribute scales.

Aiming at the safety of divers' underwater gas supply, a set of real-time monitoring and verification scheme for wearable divers' gas supply safety is designed based on underwater wireless short-distance transmission technology and MS5837 pressure sensor technology. The test shows that the scheme can efficiently collect key technical indexes such as air supply pressure and residual pressure of gas cylinders carried by divers, and achieve a wireless short-distance transmission effect of about 1.0 meters underwater, which can meet the design requirements of divers' underwater air supply safety monitoring scheme.

In the underwater acoustic physical layer key generation system, the environmental noise has a great impact on the quantization of the observation sequence. However, the quantization method based on the guard band cannot satisfy the underwater acoustic channel with random distribution, and the large transmission delay makes it difficult to employ the quantification method based on compensation. The selection quantization method based on membership is analyzed in the context of the underwater acoustic channel to reduce quantization error with the litter communication cost. The simulation results show that the proposed method can be applied to the underwater acoustic environment with serious noise and improve system performance.

Strongly robust signal feature construction is paramount to improve the accuracy of underwater acoustic (UWA) communication signal modulation recognition. But the ambient noise in the underwater channel confuses the signal features, and the single feature extraction technique poses huge difficulties to the modulation identification. In this paper, various feature extraction methods are advanced for several commonly used modulation signals in UWA communication systems. We analyze the differences in their features, and validate the advantages of multi-feature fusion in characterizing signal differences. Firstly, the time-frequency maps of different signals are gained by continuous wavelet transform (CWT) theory. Secondly, we propose the two strongest spectral lines extraction method based on power spectrum and autocorrelation spectrum, and the second strongest spectral line amplitude is used as a new and more obvious feature parameter. Thirdly, two correlation coefficient features based on circular spectral are constructed. Through analyzing the simulation data, the results show that different signal feature extraction techniques provide richer analysis domains and the multi-domain fusion features have more robust performance under signal-to-noise ratio (SNR) of 3dB.

This paper proposes a miniaturized ± 45º dual-polarized antenna, which obtains a certain bandwidth by grooving in the dipole and digging out the circular plaque at the edge, adding parasitic branches and parasitic patches at the end of the dipole to obtain a wider bandwidth, using a W-type feed, which can obtain a wider bandwidth compared to the Y-type feed, and the results show that the antenna covers the three frequency bands (3.3-5GHz) of N77, N78 and N79. And the antenna in 3.29-5.0GHz S11 and S22 are less than -15, and the gain is not less than 8dBi, the isolation between the two ports is higher than 25dB, HPBW change range within 66±8°, the size is only 0.37λ*0.37λ

Limited numbers of radio frequency (RF) test configurations have to be designed for the burn-in test, outgassing test and small signal monitoring test in vacuum for telecommunication satellite. Traditional method is to draw colored curves in schematic to recognize different RF test configuration. However, with the development of large-capacity and long-life telecommunication satellite, redundancy ring is bigger and bigger and the schematic is becoming more and more complex. Traditional drawing curve method is in-efficient and not reliable for large redundancy ring. This paper presents a novel optimum design of RF test configuration of redundancy ring for telecommunication satellite based on genetic algorithms (GA).

In order to study the propagation characteristics of ultrasonic guided waves in power plant boiler pipes and the ability to detect defects, the motion equation and dispersion equation of pipe guided waves were derived and analyzed; Numerical simulation study on the propagation process of T (0,1) mode guided waves in power plant boiler bends using finite element software. The results indicate that the longitudinal mode is more sensitive to circumferential defects, while the torsional mode is more sensitive to axial defects. When the T (0,1) mode ultrasonic guided wave propagates in a U-shaped bend, the mode of the guided wave will change.

An improved YOLOV4-tiny algorithm was proposed to address the difficulties in detecting small targets and high complexity in Printed Circuit Board (PCB) defect detection, as well as the inability to meet real-time detection requirements. Firstly, the backbone network was changed to a lighter MobileNet-V3 to solve the problem of excessive parameter quantity, making the model more lightweight. Secondly, the detection scale was increased to three, enhancing the network models to detect small target defects. An improved SPP module was proposed to further improve the feature map expression ability. Finally, the anchor box sizes were re-clustered using the K-means algorithm to accelerate network convergence. It was learned through experiments that the accuracy of this algorithm improved by 4.28%, 1.03%, and 4.94% compared to SSD, YOLOv3, and YOLOv4-tiny algorithms, respectively. The model size was reduced by 1.4 MB compared to YOLOv4-tiny, and the detection speed reached 83.33FPS that satisfies the demands for real-time detection.

Foreground detection is a significant area of study within the realm of computer vision and plays a crucial role in video-based applications. The Vibe algorithm is an efficient foreground detection method, and this paper proposes an optimized Vibe algorithm that incorporates fused edge information to address the issues of "ghosting" and noise. Initially, the edge information derived from the three-frame method is enhanced by the Canny operator. Subsequently, the OTSU algorithm is employed to filter out additional noise in the image, thereby enhancing the robustness of the algorithm. Finally, the optimized three-frame difference method is fused with the Vibe algorithm to eliminate the "ghosting" and achieve a complete foreground target. Experimental results demonstrate that the proposed algorithm effectively removes "ghosting" and noise, providing good real-time performance and robustness, thereby meeting the application requirements of real-time detection.

In this paper, a tri-band low profile antenna based on Artificial Magnetic Conductor (AMC) is designed. The antenna consists of a tri-band slot antenna and AMC reflector. The Coplanar Waveguide(CPW)-fed tri-band antenna generate multiple resonant frequencies from slots and parasitic elements. To improve the CPW antenna’s radiation performance, an AMC structure with three zero-phases of the reflection coefficient is designed and employed as the reflector. An AMC unit with three zero-phase reflection coefficients is designed and arranged in 3×3 as the reflector of the slot antenna. The distance between the slot antenna and AMC structure is 0.056λ (λ is the wavelength of 2.4 GHz in free space). Simulation results show that the antenna impedance bandwidth after loading the AMC reflector are 2.23-2.49 GHz, 5.45-5.89 GHz and 7.45-7.91GHz. Meanwhile, the peak gain are improved from 1.9 to 3.4 dBi at 2.4 GHz, 3.45 to 6.2 dBi at 5.8GHz, 3.92 to 7.6 dBi at 7.6 GHz respectively.

Terminal devices of the 1553B system interconnect with each other through the 1553B bus network. In order to insure the reliability of the communication, it is necessary to carry on a rigorous test that includes two categories of qualitative test and quantitative test on the bus network. How to test the bus network comprehensively and effectively is an important research topic, This paper will study the quantitative test of waveform dynamic characteristics in the 1553B bus network, such as zero-crossing distortion, distortion and symmetry, and propose an auto test program, and describe its hardware design, software process and some key algorisms in detail.

Point cloud registration is a key technology in point cloud processing, aiming to align two point clouds by estimating the optimal rigid transformation matrix. When the overlap rate of the point clouds to be registered is low, it is difficult for the existing point cloud registration algorithms to achieve satisfactory results. To achieve accurate registration of point clouds with low overlap rate, an ICP point cloud registration algorithm based on variable-length least trimmed squares (VLTS-ICP) is proposed, which enables the accurate registration of point clouds with low overlap rate by varying the truncated length of the LTS to make more pairs of ideal matching points be used in the current iteration for the estimation of pose transformation. It is verified that the VLTS-ICP algorithm has better accuracy and robustness, especially for low overlap point cloud registration, and its comprehensive performance is significantly better than ICP algorithm and TrICP algorithm.

With faster transmission speed and lower latency of information, V2X is gaining more and more public attention. V2X mainly interacts with vehicle-vehicle, vehicle-road, vehicle-person and so on to build real-time information map of vehicle networking. Due to cost and technology constraints, the integration of vehicle intelligence and V2X intelligence will be the main trend of the development of auto-driving. Virtual simulation can not only meet the requirements of scene coverage, but also improve the security of test. It is the main means of automatic driving verification. This paper presents a simulation verification method based on VTD for the fusion of vehicle intelligence and V2X networking. The main purpose of this method is to make up for the limitations of vehicle intelligence in perception by using V2X networking characteristics. In the simulation software, the simulation of the fusion of vehicle intelligence and V2X networking is carried out to verify the advantages and disadvantages of the auto-driving algorithm. The simulation validation method makes up for the singularity of the conventional simulation. It can verify the repeatability and security of the fusion technology several times before it is applied to the ground, and improve the test efficiency and accuracy.

Aiming at the problems of poor real-time performance and low success rate of mobile navigation information free switching caused by improper use of mobile navigation information free switching parameters, a hybrid mobile navigation information free switching method is designed by introducing virtual reality technology. The accelerated robust feature algorithm is used to identify massive navigation scenes, track the feature points of navigation scenes, and obtain 2D navigation information. On this basis, X3D technology is used to create virtual reality scene and obtain 3D navigation information. According to the obtained two-dimensional and three-dimensional navigation information, the navigation information switching process is designed, the switching parameters are defined, and the navigation information switching execution rules are specified, so as to realize the free switching of hybrid mobile navigation information in virtual reality. The simulation results show that compared with the existing hybrid mobile navigation information free switching methods, the hybrid mobile navigation information free switching method proposed in this paper greatly improves the real-time performance and success rate of switching, which fully shows that the proposed hybrid mobile navigation information free switching method has better navigation information switching effect.

As a technology that has been developed for more than ten years, small target detection has been very mature at the technical level. A large number of target detection methods such as Faster R-CNN, RetinaNet, and YOLO have emerged that can be used in the industry. However, the problem of poor detection performance of small targets has not been completely solved so far. This paper focuses on the research and analysis of the optimization of the YOLOv5 algorithm around four optimization schemes, so that the sensitivity and fineness of the YOLOv5 algorithm for small target detection have been significantly improved. The experimental results show that the CBAM module can provide more small target feature information in the feature information extraction link, thereby increasing the efficiency of the algorithm for small target detection, and the replacement of the PAN-Net structure with the Bi-FPN structure can be used in the feature information feedback link. Reduce its loss, thereby increasing the efficiency of the algorithm, and the two schemes can be well combined to further improve the detection effect of small targets.

In this paper, the radar parameters of human body junction including different behavioral physical characteristics are analyzed. The radar parameters of different activities (walking, squatting, sitting and falling) of the target under multiple observation angles are simulated. The radar characteristics of human motion such as Doppler frequency, radial distance, height and pitch angle are analyzed in depth. Finally, the most suitable view angle in the classification of each activity and the radar parameters with the most stable characteristics are further given. The feature extraction method presented in this paper has an explicit physical interpretation and good angular adaptability. Neural networks combined with the proposed feature extraction method can have better interpretation of the activity classification results.