Abstract:The system emulator creates a virtual environment by emulating hardware resources such as processor, memory, and peripherals, which can support software running and debugging of different architectures and greatly shorten the cross-architecture software development cycle. The emulator usually supports instruction tracing and can be employed for analysis by recording the instruction sequence of program running, such as running time evaluation and behavior pattern analysis related to the program, and joint emulation of software and hardware. As the mainstream emulators supporting RISC-V architecture, both QEMU and Spike support instruction tracing. However, they are time- and space-expensive and inefficient when dealing with large-scale applications. Thus, this study proposes an instruction tracing technology with QEMU. When instructions are traced without distortion, static information such as basic blocks and control flow charts in the program is decoupled from branch selection and other dynamic information. Compared with the native instruction tracing implemented by QEMU, the proposed technology reduces the time overhead by more than 80% and the space overhead by more than 95%. Additionally, based on RISC-V architecture, this study realizes off-line analysis of instruction sequences in various scenarios, such as instruction classification statistics, program hotspot marking, and program behavior analysis.

Abstract:Traditional x86-based and software-based user-mode memory safety defenses can hardly be deployed in a production-ready environment due to significant runtime overheads. In recent years, as mainstream commercial processors begin to provide hardware security extensions and open-source architectures like RISC-V rise, hardware-assisted memory safety protections have become popular, and their implementations are based on various architectures, such as x86-64, ARM, and RISC-V. This study discusses user-mode memory safety defenses on the RISC-V architecture and compares the features of x86-64, ARM, and RISC-V in the context of security defense design. RISC-V has some advantages over other architectures due to its opening ecosystem, making the implementation of some low-cost and promising defense techniques possible.

Abstract:RISC-V instruction set architecture (ISA) has promoted the rapid development of the RISC-V hardware platforms, leading to growing demands for efficient and easy-to-use operating systems running on RISC-V architecture. As a distributed open-source mobile operating system, OpenHarmony continues to evolve with constant prosperous ecology. However, adapting OpenHarmony to RISC-V ISA platforms poses new challenges, including software stack and chip porting. This study presents an approach and methodology for porting the OpenHarmony standard system to the RISC-V QEMU platform. Based on adapting critical software stack components and porting the graphics display driver on the QEMU RISC-V virtualization hardware platform, the OpenHarmony standard system successfully starts on the QEMU RISC-V virtualization hardware platform and enters the system desktop. This achievement provides developers with a platform to test and apply the OpenHarmony standard system on RISC-V platforms and serves as a reference for porting the OpenHarmony standard system to new RISC-V hardware platforms.

Abstract:As a lightweight standard C library, musl libc features a small code base providing comprehensive POSIX interface support, and high portability and support for various architectures and operating systems. It is widely employed in embedded systems, Web servers, containers, and other fields. RISC-V instruction set is an open source instruction set that has released the relatively stable SIMD instruction set at present. Meanwhile, the RISC-V ecological software environment has ushered in a new optimization boom, but the RVV extension optimization of the musl libc library is still a research gap. Based on the collaborative research of the musl libc basic library and RISC-V RVV extended instruction set, this study proposes an implementation scheme compatible with the basic instruction set and vector extended instruction set. The common C library functions strlen and memset are optimized by the vector extended instruction set, and comparative analysis is carried out on gem5 simulator. The experimental results show that compared with the implementation of C language, the performance of strlen function optimized by RVV is improved by 83%–703% on average, and that of memset function is improved by 85%–334% on average.

Abstract:Deep learning models require certain interpretability in practical applications in certain scenarios, and vision is a basic tool for humans to understand the surrounding world. Visualization technology can transform the model training process from an invisible black box to an interactive and analyzable visual process, effectively improving the credibility and interpretability of the model. At present, there is a lack of review on deep learning model visualization tools in related fields, as well as a lack of research on the actual needs of different users and the evaluation of user experience. Therefore, this study summarizes the current situation of the application of visualization tools in different fields by investigating the literature related to interpretability and visualization in recent years. It proposes a classification method and basis for target user-oriented visualization tools and introduces and compares each type of tool from the aspects of visualization content, computational cost, etc., so that different users can select and deploy suitable tools. Finally, on this basis, the problems in the field of visualization are discussed and its prospects are provided.

Abstract:At present, the yak breeding method in the Qinghai-Tibet Plateau region of China is mainly based on traditional manual grazing. To solve the problem that human breeding methods cannot quickly track and count the number of yaks, an improved YOLOv5 and Bytetrack yak tracking method is proposed in this study to achieve the fast detection and tracking of yaks under video input. The YOLOv5 object detection network based on deep learning, combined with optimization methods such as coordinate attention, cross-scale feature fusion, and atrous spatial pyramid pooling pyramid, is adopted to reduce the difficulty of detection and misdetection caused by occlusion in yak detection, so as to accurately detect yak targets in videos. The Bytetrack tracker is used to implement the inter-frame object association through Kalman filtering and Hungarian algorithm, and the IDs are matched to the targets. The model is trained by using part of the yak data in ImageNet Dataset and yak sample images collected from the Yushu region of Qinghai. The experimental results show that the average detection accuracy of the improved model proposed in this study is 98.7%, which is 1.1, 1.89, 8.33, and 0.4 percentage points higher than the original YOLOv5s, SSD, YOLOX, and Faster RCNN models, respectively. It can converge quickly and has the best detection performance. The improved YOLOv5s and Bytetrack tracking results are the best, with MOTA increased by 7.1646%. The improved model developed in this study can detect and track yaks more quickly and accurately, providing technical support for the intelligent development of animal husbandry in the Qinghai region.

Abstract:As Web applications become increasingly complex, their security issues happen frequently. Web application security testing has become one of the research priorities in the field of software testing. Vulnerability reports aim to document Web application security issues and assist Web application testing to improve its security and quality. However, how to automatically identify the key information in vulnerability reports and reproduce the vulnerability is still a research challenge. To this end, this study proposes an automatic approach to comprehend vulnerability reports and reproduce the vulnerability. Firstly, based on the characteristics of vulnerability reports, the study summarizes their grammar dependency patterns and combines them with dependency syntactic parsing techniques to parse vulnerability descriptions and extract key information about vulnerability triggers. Secondly, unlike conventional natural language descriptions, the payload of Web vulnerability is usually an illegal string, mostly in the form of a code fragment. For this reason, the study designs extraction rules for the payload solely to improve the extraction of vulnerability reports. On this basis, considering that vulnerability reports and Web application text descriptions are different but semantically similar, the study proposes a semantic similarity-based method to achieve the automatic reproduction of Web application vulnerability. To verify the effectiveness of this study, 400 vulnerability reports are collected from more than 300 Web application projects in the vulnerability collection platform Exploit-db, and their grammar dependency patterns are summarized. A total of 26 real vulnerability reports involving 23 open-source Web applications are used for vulnerability reproduction experiments. The results show that the proposed method can effectively extract key information from vulnerability reports and generate feasible test scripts to reproduce vulnerability, reducing manual operations, and improving the efficiency of vulnerability reproduction.

Abstract:The problem of imbalanced datasets has attracted people’s attention since two decades ago, and various solutions have been proposed. Mixup is a popular data synthesis method in recent years, with many variants extended. However, there are not many Mixup variants proposed for imbalanced datasets. This study proposes a Mixup variant, namely Borderline-mixup, to address the classification problem of imbalanced datasets, which uses a support vector machine (SVM) to select boundary samples and increases the probability that the boundary sample is sampled in the sampler. Two boundary samplers are constructed to replace the original random sampler. Extensive experiments have been conducted on 14 UCI datasets and CIFAR10 long-tail datasets. The results show that Borderline-mixup has outperformed Mixup consistently on UCI datasets by up to 49.3% and on CIFAR10 long-tail datasets by about 3%–3.6%. Therefore, the proposed Borderline-mixup is effective in the classification of imbalanced datasets.

Abstract:In this study, a mathematical model aiming at minimizing the total distribution distance is established for the vehicle routing problem of simultaneous delivery and pickup with time window constraints. According to the characteristics of the model, a discrete grey wolf optimization (DGWO) algorithm is proposed to solve the problem on the basis of preserving the search mechanism of the grey wolf optimization (GWO) algorithm. Multiple strategies are adopted to construct the initial solution of the population, and the unfeasible solution is allowed to expand the search area of the population; the neighborhood search strategy with scoring strategy is introduced to adjust the probability of each operator so that the algorithm can select the operator with better optimization effect; the deletion-insertion mechanism is used to explore the high-quality solution region and accelerate the convergence of the population. The standard data set is tested in the simulation experiment, and the experimental results are compared with the p-SA algorithm, DCS algorithm, VNS-BSTS algorithm, and SA-ALNS algorithm. The experiment shows that the DGWO algorithm can effectively solve the vehicle routing problem of simultaneous delivery and pickup with time window constraints.

Abstract:In automated manufacturing systems (AMSs), deadlock is an urgent problem to be solved, which is mainly caused by circular waiting for resources. To solve this problem, this study first builds special resource marked graphs (SRMGs) based on the characteristics of resource-oriented Petri nets (ROPNs). Secondly, the relationship between a deadlock and the saturated circuit is established in SRMGs. Unsafe markings can be prevented by adding controllers to some special circuits. Next, considering the problem of resource failure, the resource buffer subnet is added to the hazardous place. This ensures that parts requiring failed resources do not block the continuous production of other parts. Compared with existing controllers, each controller in this study has a control switch that allows more safe markings to occur by changing the capacity of the control place in real time.

Abstract:To solve the problems of lengthy paragraphs, sparse data, scattered knowledge, and poor specification of text data in psychological medicine, a method based on the pre-trained model of multi-level feature extraction capability (MFE-BERT) and forward neural network attention (FNNAttention) mechanism is proposed for the construction of psychomedical knowledge graphs. Based on the BERT model, MFE-BERT merges and outputs all the internal encoder layer features to obtain feature vectors with more semantics. At the same time, the FNNAttention mechanism is applied to the two composite models to strengthen the word-level relationship and solve the semantic dilution of long text paragraphs. In the self-created psychomedical datasets, the compound neural network models of MFE-BERT-BiLSTM-FNNAttention-CRF and MFE-BERT-CNN-FNNAttention are designed for psychomedical entity recognition and entity relationship extraction respectively. The entity recognition F1 value reaches 93.91% and the entity relation extraction precision rate reaches 89.29%. The entity alignment is carried out by merging text similarity and semantic similarity. The collated data are stored in a Neo4j graph database, and a psychomedical knowledge graph containing 3652 entities and 2396 relationships is constructed. The experimental results show that it is practical and feasible to construct a psychomedical knowledge graph based on the MFE-BERT model and the FNNAttention mechanism, and the psychomedical knowledge graph built by the proposed improved models can be better applied in psychomedical information management, providing a reference for psychomedical data analysis.

Abstract:With the rapid development of smart stations and cloud computing, the deployment of large-scale video surveillance systems for pedestrian detection in subway stations is becoming more and more important, which plays an important role in passenger flow monitoring, passenger guidance, and behavior warning. In practical engineering applications, a lightweight pedestrian detection algorithm MCA-YOLOv5s is proposed due to the adverse effects of limited computing resources and difficult samples caused by multi-scale and multi-angle occlusion. Firstly, MobileNetv3 replaces the YOLOv5 backbone network to achieve lightweight network model processing, and PConv replaces DWConv in the MobileNetv3 network to reduce redundant computation and memory access. Secondly, the coordinate attention mechanism is incorporated in the C3 module of the feature fusion stage to make the model pay more attention to pedestrian position information. At the same time, the loss function CIoU is replaced by Alpha IoU to increase the weight of the High Loss target and the regression accuracy of the bounding box. Finally, the improved network model is compressed by FPGM pruning to improve the loading and running speed of the model. The improved model is deployed in Huawei Atlas 300 AI accelerator to detect pedestrians in subway stations. The average accuracy is 94.1%, and the detection speed is 104.1 fps. The actual engineering practice shows that the detection speed of the improved algorithm is increased by 71.8%, saving the hardware deployment resources in the station and meeting the actual engineering needs of pedestrian monitoring and management in subway stations with large passenger flow.

Abstract:In this study, a non-touch online signature authentication system based on the MediaPipe framework is designed and implemented for the reliability and device dependency of handwritten signature authentication. The system utilizes MediaPipe as the underlying framework, captures the online handwritten signatures through video, extracts the temporal features of signature trajectory points as matching templates, and constructs a signature authentication model using a weighted joint probability strategy. The model achieves an average equal error rate (EER) of 3.04% on edge devices. An application designed based on PyQt is used as a visual UI interface to enable online non-touch signature authentication in video scenarios. This system uses real-time video sensing interaction to achieve online signature authentication without the need for other external devices, resulting in lower device dependency and greater authentication reliability.

Abstract:In recent years, the recommendation system has become a hot spot in the field of data analysis and mining, as well as information retrieval. However, there are still problems in some recommendation systems serving the multi-interest preferences of users. Firstly, the users’ interests are not single, and the preference for multiple interests is not equal. Secondly, it is not sure whether the users’ current interests will continue in the future. Therefore, this study proposes a MIES algorithm model by utilizing the items that users participate in to generate multiple interests and capture the sustainability of their personalized interests. The model effectively captures users’ diverse latent interests while emphasizing the sustainability of their interests, thus improving the quality of recommendations. Comparative experiments demonstrate that the model effectively addresses the challenges of capturing users’ multidimensional interests in recommendation systems and ensuring the sustainability of personalized interests.

Abstract:Compared with traditional methods, current deep learning-based image completion methods have achieved better repair results. However, most of them overlook the establishment of pixel long-distance dependence, and deep learning models have poor performance in dealing with large irregular missing areas, resulting in insufficient overall fit of the generated image. On the other hand, many completion algorithms that retain more detailed information by fusing multi-scale receptive fields are affected by changes in the input scale and the completion target scale as they cannot adjust the receptive field dynamically, resulting in significant artifact errors in the generated results. In response to such problems, this study proposes a completion algorithm based on fast Fourier transform and selective convolutional kernel network, which ensures the efficient operation of the model while achieving pixel long-distance dependence. In addition, this algorithm also improves the selective convolutional kernel network, which can adaptively adjust the corresponding weights according to the contribution of each convolutional kernel feature. It provides accurate local information supplementation for the model, ultimately generating completion results with higher global fusion and richer local details. The experiments on the Celeb-A and Place2 datasets show that the proposed method not only surpasses existing cutting-edge image completion methods in PSNR and SSIM metrics but also has significant advantages in processing images with occlusion rates of over 80%, which can generate more realistic results.

Abstract:In view of the problem of incomplete feature extraction and large model size caused by ignoring global features in shallow networks in existing pneumonia medical image recognition research, a lightweight model based on convolutional neural network (CNN) and attention mechanism is proposed to improve the recognition efficiency of pneumonia types. A lightweight model structure is used to reduce the number of model parameters. By increasing the convolution kernel, efficient channel attention and self-attention mechanisms are introduced to solve the problem of loss of important network information and the inability to extract underlying global information. Local and global information is extracted in parallel through dual branches, and multi-scale channel attention is utilized to improve the fusion quality of the two. The CLAHE algorithm is employed to optimize the original data. The experimental results show that the accuracy, sensitivity, and specificity of the model are increased by 2.59%, 3.1%, and 1.38% respectively compared with those of the original model while ensuring lightness, and the proposed model outperforms other current excellent classification models and has stronger practicability.

Abstract:As urbanization accelerates and the population continuously increases, the utilization and management of land resources have become increasingly important. The development of high-resolution remote sensing technology provides a new approach for detecting land cover changes. Currently, most remote sensing image change detection tasks mainly focus on detecting significant changes in buildings, and there is a lack of research on detecting changes in land cover categories. In this study, based on a public dataset, more land cover change scenarios are annotated. Combining the original semantic segmentation backbone network with a Siamese network structure, this study proposes a detection model suitable for tasks of detecting changes in land cover categories. The model incorporates a change guidance module in the feature extraction stage to assist the network in focusing on change information in the two temporal images. A channel information interaction module is added at different stages of the network to enhance the fusion of information from different feature maps. Additionally, a feature alignment module is added to the last layer of the feature extraction stage to alleviate feature offset caused by downsampling. Experimental results on a dataset of detecting changes in land cover categories demonstrate that the proposed method can effectively extract change information from the image and improve the segmentation accuracy.

Abstract:Federated learning allows multiple users to collaboratively train models without sharing the original data. To ensure that users’ local datasets are not leaked, the existing works propose secure aggregation protocols. However, most of the existing schemes fail to consider global model privacy, and the system is at a high cost of computational and communicational resources. In response to the above problems, this study proposes an efficient and secure privacy-preserving secure aggregation scheme for federated learning. The scheme uses symmetric homomorphic encryption to protect the privacy of the user model and the global model and adopts secret sharing to solve users’ dropout. At the same time, the Pedersen commitment is applied to verify the correctness of the aggregation results returned by the cloud server, and the BLS signature is utilized to protect the data integrity during the interaction between the users and the cloud server. In addition, security analysis illustrates that the proposed protocol is of provable security; performance analysis indicates that the protocol is efficient and practical for federated learning systems with large-scale users.

Abstract:A new algorithm based on the bald eagle search algorithm (NBES) is proposed to address the drawbacks of poor stability and low accuracy of the solution and poor robustness of the bald eagle search (BES) algorithm. First, the sine cosine optimization mechanism algorithm is fused in the search space selection stage of the BES algorithm, and the fused position update formula is constructed. Secondly, the inertial weight adaptive position update strategy is added in the search space prey phase of the BES algorithm. Finally, the position update formula is redefined by fusing the firefly optimization mechanism algorithm in the swoop phase of the BES algorithm. The performance of the NBES algorithm is verified by 11 standard test functions, and the experiments show that the NBES algorithm outperforms the BES algorithm in terms of search accuracy, convergence speed, and robustness. To verify the practical application value of the new algorithm, the hyperparameter learning rate in the convolutional neural network (CNN) is optimized by using the NBES algorithm, and the optimized image classification model is used in medical image pathology classification prediction, and the experiments show that the classification accuracy of the optimized CNN model is improved by 9%.

Abstract:The samples to be tested for metal surface defects are often characterized by low resolution, fuzzy defect boundaries, dense defects, and small defect targets. At the same time, the constructed detection model has a large number of hyperparameters that need to be manually adjusted and lacks the adaptive parameter adjustment ability. In this study, a surface defect super-resolution detection algorithm based on Bayesian optimization is proposed. Through the design of fine layered structure, the receptive field of the backbone network feature map is enriched; the extraction of high-low frequency information is enhanced; the high-resolution image with clear edge texture is reconstructed. By constructing the bottleneck residual dense structure, the shallow and deep features of the backbone feature extraction network are enriched, and the classification ability and the localization ability of the model for small targets and dense targets are improved. The key hyperparameters of the detection model are optimized adaptively by a Bayesian optimization algorithm with low time cost. Experiments show that mAP_0.5 for six types of metal surface defects in the NEU-DET dataset can reach 0.782, and the detection speed can reach 102 f/s, which is superior to other detection algorithms.

Abstract:Unmanned aerial vehicles (UAVs) can act as air edge servers to provide services for ground mobile terminals in disaster areas where earthquakes, typhoons, floods, and mudslides have caused severe damage. However, it is difficult to complete complex computationally intensive tasks in real time due to the limited computation and storage capacity of a single UAV. In this study, a multi-UAV-assisted mobile edge computing model is first investigated and a mathematical model is built. Then a partially observable Markov decision process is established and an improved multi-agent deep deterministic policy gradient (MADDPG) algorithm based on the composite priority experiential replay sampling method (CoP-MADDPG) is proposed to jointly optimize time delay, energy consumption, and flight trajectory of UAVs. Finally, the simulation experimental results show that the proposed algorithm outperforms other benchmark algorithms in terms of total reward convergence speed and convergence value, and can provide services for about 90% of ground mobile terminals, proving the effectiveness and practicality of the proposed algorithm.

Abstract:Massive trajectory data pose challenges to management analysis and data mining, and trajectory compression technology has become an effective solution to this problem. Aiming at the problem that most current trajectory compression algorithms need human intervention to set thresholds, this study proposes an adaptive trajectory inflection point extraction and compression algorithm which combines the idea of feature clustering and trajectory partition. Based on the global and local direction characteristics of the trajectory, the algorithm carries out the rough trajectory division, sub-trajectory merging, and fine trajectory division. The experimental results show that with the increasing trajectory size, the proposed algorithm can produce lower direction error and maintain a higher compression rate than other algorithms. The algorithm features adaptive and high-precision inflection point recognition and still has a high reference value under other trajectory compression scenarios.

Abstract:As the population size of the micro-population teaching and learning optimization algorithm is small, it is hard to maintain its population diversity. To improve the search performance of the micro-population teaching-learning-based optimization algorithm, a micro-population teaching-learning-based optimization algorithm based on multi-source gene learning (MTLBO-MGL) is proposed. In MTLBO-MGL, the teaching stage and the learning stage are used to evolve individuals at the gene level via the random selection strategy. Moreover, the population diversity is detected at the gene level and the sparse spectral clustering is utilized to cluster the population on each dimension. Different evolutionary strategies are selected to improve the search performance of the proposed algorithm based on the diversity detection result and the clustering result. The performance of the proposed algorithm is compared with the other four micro-population evolutionary algorithms on 28 test functions. The simulation results prove that the overall performance of the proposed algorithm is significantly better than the other four compared algorithms. The proposed algorithm is also applied to solve the UAV 3D path planning problem, and the results show that MTLBO-MGL can achieve better results on this scenario.

Abstract:The current sentiment classification methods often ignore the relative positional features between different words, which makes it difficult for the model to learn the best positional representation of words. To solve this problem, a sentiment classification algorithm based on Gaussian distribution guided position relevance weight is proposed. First, the positional relevance between each word and other words is calculated. Second, the positional relevance is modeled by using an improved Gaussian distribution function, and the results are multiplied with the feature vectors of the words to generate a positional-aware representation of the words. Finally, the algorithm is integrated into the traditional model to verify its effectiveness. The experimental results show that the proposed method obtains higher accuracy than the traditional model, with improvements of 2.98%, 5.02%, and 10.55% in terms of in-domain, out-of-domain, and adversarial evaluation metrics, respectively, indicating its excellent practical value.

Abstract:Existing image restoration methods generally suffer from structural misalignment and blurred edges of the restored area, which is due to the under-utilization of structural information in known areas during image restoration. To this end, a color clustering-based image restoration algorithm with an encoder-decoder structure is proposed in this study. The algorithm uses a progressive image restoration network structure, taking the results of the images after color clustering as input, and the images processed by the clustering algorithm better preserve the structural information. Therefore, the structural information can be fully utilized in the subsequent image texture restoration process. Meanwhile, to further reduce the computational overhead of the network, a cross-attention mechanism is introduced to model the global dependence of images from both horizontal and vertical dimensions. The experimental results show that the image restoration algorithm proposed in this study can effectively alleviate the blurring of the edges in the restored areas, and compared with several mainstream image restoration algorithms, the proposed image restoration algorithm can produce more realistic output results in the case of large missing areas.

Abstract:Traditional clustering algorithms can split the dataset into different clusters, whereas these clusters are usually difficult to explain. Iterative mistake minimization (IMM) is a common explainable clustering algorithm, which constructs a threshold tree from a single feature, and each cluster can be explained by feature-threshold pairs on the path from the root node to the leaf node. However, the threshold tree only considers the feature-threshold pair with the fewest errors when dividing the data in each round, and this greedy method is easy to lead to the local optimal solution. To solve this problem, this study introduces beam search, which slows local optimization by retaining a predetermined number of states in each round of division, thereby improving the consistency between the clustering provided by the threshold tree and the initial clustering. Finally, the effectiveness of the algorithm is verified by experiments.

Abstract:As a fundamental component of a software system, the kernel of an operating system plays a crucial role in constructing a highly trusted software system. However, in practical verification, there are still many difficulties in invariant definition of global properties, and formal description and verification of complex data structure programs in the kernel of an operating system. Given the global properties satisfied by the kernel of an operating system, this study defines these properties at the code level on a function-by-function basis through global invariants and conducts formal verification in different functions to prove that each function conforms to the global properties of the operating system kernel. To formalize the complex data structure programs frequently adopted in the kernel of an operating system, the study proposes a method employing nested shape graph logic by extending the shape graph theory and provides a correctness proof for this method. Finally, it verifies the code related to task creation and scheduling, and message queue creation and operation in the operating system kernel.

Abstract:Mobile edge computing (MEC) has gradually become an effective means of alleviating data overload. However, in highly crowded scenarios, edge servers fixed on base stations may fail to provide efficient services due to network overload. In view of the communication demands of low latency, a dual-layer unmanned aerial vehicle (UAV) with high mobility and easy deployment becomes an ideal choice for task offloading. The top UAV (T-UAV) equipped with computing resources can provide offloading services for the bottom UAV (B-UAV) capturing the on-site scene. The B-UAV equipped with a shooting device can choose to perform local computing or partially offload tasks to T-UAV for computation. In this study, a dual-layer UAV-assisted MEC system model is constructed, and a new offloading algorithm named deep deterministic policy gradient offloading algorithm based on composite prioritized experience replay (DDPG-CPER) is proposed. The algorithm comprehensively considers the continuity of decision variables and optimizes the task execution latency under constraints such as T-UAV resource scheduling and mobility, thus improving processing efficiency and response speed, so as to ensure a real-time viewing experience for on-site spectators. The simulation results show that the proposed algorithm exhibits faster convergence speed than baseline algorithms such as DDPG and can significantly reduce processing latency.

Abstract:This study proposes a multi-range instrument recognition method based on YOLOv7+U2-Net to address the difficulties in locating instruments and low inference accuracy in the detection and recognition process of pointer instruments in complex environments. In order to improve the input image quality of the U2-Net model, a YOLOv7 detector with high inference accuracy and speed is selected. The detected and cropped images are used as the input image dataset of the model. At the same time, rotation correction is applied to the input image, making the model suitable for multi-angle instrument recognition. In response to issues such as poor accuracy and slow inference speed of instrument readings, the ordinary convolution of RSU4-RSU7 in the U2-Net decoding stage has been replaced with deep separable convolution. On this basis, an Attention mechanism has been introduced to accelerate the overall inference speed and accuracy. In addition, in order to improve the universal applicability of this method, a recognition accuracy discrimination method within multiple threshold ranges is proposed to adapt to various application scenarios. Through comparative experiments, it has been shown that when evaluated on the collected dataset, compared with template matching, SegNet, PSPNet, Deeplabv3+, and U-Net methods, the proposed method achieves a recognition accuracy of 96.5% and performs well in multiple threshold ranges.

Abstract:The advance in network technology and the rise of multi-access edge computing have led to the deployment of computation and network resources closer to the end users. As the service numbers increase, it is a challenge to predict the quality of service (QoS) in real-time and accurately in the complex and dynamic edge computing environment to better recommend services to users. In this study, a deep neural model for real-time QoS prediction based on service load (QPSL) is proposed, which can provide missing load condition awareness and cycle awareness for QoS prediction in edge computing. Firstly, the service load condition is characterized, and the features of the time-series are obtained by the time-series decomposition module. Secondly, CNN and BiLSTM are combined to learn the potential time-series relationships and generate the state vectors at different time intervals. Then, the state vectors at future time intervals are constructed by assigning weights to the historical state vectors based on the Attention mechanism. Finally, contextual embedding vectors and state vectors are fed into the perception layer to complete the real-time QoS prediction. Extensive experiments are conducted based on a real fusion dataset, and the results show that QPSL improves MAE by 10.28% and 10.87% on average for response time and throughput tasks respectively, outperforming existing time-aware QoS prediction methods.

Abstract:The prediction process of traditional time series prediction methods cannot push out the sharing mode on the same data set, while machine learning methods fail to handle nonlinear and large-scale data sets well, and feature engineering needs to be designed manually. The deep learning method makes up for the disadvantage of the traditional prediction method that requires high computation and high manpower, and it uses automatic learning feature engineering instead of manually designed feature engineering. However, prediction methods that only use deep learning make fewer structural assumptions and typically require higher computational resources and a large amount of data to learn accurate models. In response to the above issues, this study proposes to use empirical mode decomposition (EMD) fusing t-tests to divide the sequence into high-frequency and low-frequency components and further process the data using traditional standard template library (STL) sequence decomposition methods for high-frequency components. The high-frequency and low-frequency components are predicted separately by Prophet. The experimental results show that compared with traditional long short-term memory (LSTM) network and Prophet prediction models, the periodic data decomposed by STL sequence can improve the overall prediction accuracy of the model, while the Prophet model fused with EMD greatly improves training efficiency.

Abstract:Rock microscopic image stitching is a key part of rock analysis and research. The rock microscopic images are large in number (hundreds of images) and rich in content and contain many similar and confusing areas, which result in low stitching efficiency and low alignment accuracy. In addition, the stitching of multiple images will result in error accumulation and misalignment. For this problem, a similar region-SURF (SR-SURF) method for rock microscopic image stitching is proposed. Firstly, similar regions are quickly extracted by using hash fingerprints. Secondly, feature points are detected in this region. Then the improved random sample consensus (RANSAC) algorithm is used to eliminate the wrong matching points. The misaligned image is corrected by the best matching template. Finally, the least squares method is introduced to eliminate the cumulative error caused by the cumulative multiplication of homography matrices. The experimental results show that the algorithm proposed in this study eliminates the cumulative error caused by multiple image stitching and solves the problem of stitching misalignment, which improves the stitching speed and alignment accuracy. It has high practical value and promotes the digital storage process of rock slices.