Abstract:Long-term object tracking remains a formidable challenge compared to short-term object tracking. However, existing long-term tracking algorithms often perform poorly when faced with challenges such as targets frequently appearing and disappearing, and drastic changes in target appearance. This study proposes a novel, robust, and real-time long-term tracking framework based on local search modules and global search tracking modules. The local search module utilizes the TransT short-term tracker to generate a series of candidate boxes, and the best candidate box is determined through confidence scoring. A novel global search tracking module is developed for global re-detection, based on the Faster R-CNN model, with the introduction of Non-Local operations and multi-level instance feature fusion modules in the RPN and R-CNN stages, aiming to fully exploit target instance-level features. To improve the performance of the global search tracking module, a dual-template update strategy is designed to enhance the robustness of the tracker. By utilizing templates updated at different time points, the tracker can better adapt to target changes. The target presence is determined based on local or global confidence scores, and the local or global search tracking strategy is selected in the next frame. Additionally, the local search module is capable of estimating the position and size of the target. Moreover, a ranking loss function is introduced for the global search tracker, implicitly learning the similarity between region proposals and the original query target. A large number of experiments are conducted on multiple tracking datasets to comprehensively assess the proposed tracking framework. The results consistently demonstrate that the proposed tracking framework achieves satisfactory performance.

Abstract:The uncertainty of neural networks reflects the predictive confidence of deep learning models, enabling timely human intervention in unreliable decision-making, which is crucial for enhancing system safety. However, existing measurement methods often require significant modifications to the model or training process, leading to high implementation complexity. To address this, this study proposes an uncertainty measurement approach utilizing neuron statistical modeling and analysis with activation values within a single forward propagation. An improved kernel density estimation technology is employed to construct neuron activation distributions and stimulate neuron normal operating range. Subsequently, a neighborhood-weighted density estimation method is utilized to calculate anomaly factors, effectively qualifying deviations of test samples from neuron activation distribution. Finally, by statistically combining the anomaly factors of each neuron, the cumulative anomaly factors of the sample provide a new perspective in assessing model uncertainty. Experimental results across multiple public datasets and models visually demonstrate the significant effectiveness of the proposed method in distinguishing between in-domain and out-of-domain samples through visualizing feature maps. Moreover, the method exhibits exceptional performance in out-of-domain detection tasks, with AUROC exceeding other methods across various experimental setups, validating its generality and effectiveness.

Abstract:Hierarchical federated learning (HFL) aims to optimize model performance and maintain data privacy through multi-layered collaborative learning. However, its effectiveness relies on effective incentive mechanisms for participating parties and strategies to address information asymmetry. To address these issues, this study proposes a layered incentive mechanism for protecting the privacy of end devices, edge servers, and cloud servers. At the edge-device layer, edge servers act as intermediaries, using the multi-dimensional contract theory to design a variety of contract items. This encourages end devices to participate in HFL using local data without disclosing the costs of data collection, model training, and model transmission. At the cloud-edge layer, the Stackelberg game models the relationship between unit data reward and data size between a cloud server and edge servers and subsequently transforms it into a Markov process, all while maintaining the confidentiality of the edge servers’ unit profit. Then, multi-agent deep reinforcement learning (MADRL) is used to incrementally approach the Stackelberg equilibrium (SE) while ensuring privacy. Experimental results indicate that the proposed incentive mechanism outperforms traditional approaches, yielding an almost 11% increase in cloud server revenue and an approximately 18 times improvement in the cost-effectiveness gained.

Abstract:The current image denoising algorithms based on deep learning are unable to consider the local and global feature information comprehensively, which in turn affects the image denoising effect at the details. To address this problem, this study proposes a hybrid CNN and Transformer image denoising network (HCT-Net). First, CNN and Transformer coupling block (CTB) is proposed to construct a two-branch structure that integrates convolution and channel self-attention to alleviate the high computational overhead caused by relying solely on the Transformer. At the same time, the attention weights are dynamically allocated so that the network focuses on important feature information. Secondly, the self-attention enhanced convolution module (SAConv) is designed to adopt the progressive combination of modules and nonlinear transformations to attenuate the noise signal interference and identify local features under complex noise levels. Experimental results on six benchmark datasets show that HCT-Net has better feature perception ability than some current advanced denoising methods and can suppress high-frequency noise signals to recover the edge and detail information of images.

Abstract:As a very challenging project in target detection, small target detection is widely distributed in daily life. In video surveillance scenarios, pedestrians’ faces about 20 meters away from the camera can be considered small targets. Due to the possibility of mutual occlusion of faces and their susceptibility to noise and weather, lighting conditions, the performance of existing target detection models on such small targets is inferior to that on medium and large targets. To address these issues, this study proposes an improved YOLOv7 model with a high-resolution detection head and transforms the backbone network based on GhostNetV2. At the same time, the PANet structure is replaced by the BiFPN and SA attention modules combined to enhance the multi-scale feature fusion capability; the original CIoU loss function is improved by combining the Wasserstein distance, reducing the sensitivity of small targets to anchor frame position offset. This study conducts comparative experiments on the public dataset VisDrone2019 and a self-made video surveillance dataset. Results show that the mAP of the improved method proposed in this study improved to 50.1% on the VisDrone2019 dataset and is 1.6 percentage points higher than existing methods on the self-made video surveillance dataset, which effectively improves the ability of small target detection and achieves good real-time performance on the GTX1080Ti.

Abstract:In semantic segmentation tasks, the downsampling process of the encoder can lead to a decrease in resolution, resulting in the loss of spatial information details in the image. As a result, segmentation discontinuity or incorrect segmentation may occur at object edges, which can damage overall segmentation performance. To address the above issues, an image semantic segmentation model EASSNet based on edge features and attention mechanisms is proposed. Firstly, the edge detection operator is used to calculate the edge map of the original image, and edge features are extracted through pooling downsampling and convolution operations. Next, edge features are fused into deep semantic features extracted by the encoder, restoring the spatial detail information of downsampled feature images, and strengthening meaningful information through attention mechanisms to improve the accuracy of object edge segmentation and overall semantic segmentation performance. Finally, EASSNet achieves the average intersection over the union of 85.9% and 76.7% on the PASCAL VOC 2012 and Cityscapes datasets, respectively. Compared with current popular semantic segmentation networks, EASSNet has significant advantages in overall segmentation performance and object edge segmentation.

Abstract:Multi-task learning is widely used in the field of natural language processing, but multi-task models tend to be sensitive to the relevance between tasks. If the task relevance is low or the information transfer is unreasonable, the task performance may be seriously affected. This study proposes a new shared-private structure multi-task learning model, BERT-BiLSTM multi-task learning (BB-MTL). It designs a special parameter optimization method, meta-learning-like train methods (MLL-TM) for the model with the help of meta-learning ideas. Further, a new information fusion gate, Softmax weighted linear gate (SoWLG), is introduced for selectively fusing the shared and private features of each task. To validate the proposed multi-task learning method, a series of experiments are conducted by combining the tasks of hate-speech detection, personality detection, and emotion detection, taking into account the fact that user behavior on the Internet is closely related to individual characteristics. The experimental results show that BB-MTL can effectively learn feature information in relevant tasks, and the accuracy rates reach 81.56%, 77.09%, and 70.82% in the three tasks, respectively.

Abstract:Recently, security issues such as identity authentication and digital signatures in industrial control systems have received more and more attention. This study introduces the decentralized certificateless (CFL) cryptography authentication system into the identity authentication of the industrial control system and proposes a signature authentication scheme for the industrial control system based on CFL. It builds the CFL-SYS authentication model for the industrial control system based on the CFL authentication system and introduces UKey as the certificated carrier to decentralize the signature verification process. A random private key and a flagging private key are generated by calculating the hash value of the user ID to realize one-person-one-key encryption, which satisfies the user’s private ownership of the private keys and protects the user’s privacy. Theoretical analysis and experimental results show that the proposed scheme can meet the millisecond-level application requirements in terms of throughput and system verification response time, and can provide an autonomous, reliable, and efficient signature authentication scheme for large-scale industrial control systems.

Abstract:Image segmentation has gradually developed from traditional threshold-based methods to convolutional neural network (CNN)-based methods. Traditional CNNs are outstanding in the field of segmentation, but the limitations of slow training speed and low segmentation accuracy are gradually emerging. To overcome these limitations, this study proposes an image segmentation recognition method based on the BM-TransUNet network, which is an improvement. A depth-separable convolution module is added to the first layer of the TransUNet network, and an attention mechanism module is introduced to the convolution layer of the encoder under-sampling so that the algorithm can better explore the features of the segmented objects. At the same time, a multi-scale feature fusion module, the feature pyramid network (FPN), is introduced between the decoder and encoder. In this study, a self-made posterior pharyngeal wall dataset is used for image segmentation training, and the effects of the trained BM-TransUNet network are compared with various traditional segmentation networks. Experimental results show that, compared to other traditional deep learning models, the identification method of the BM-TransUNet network exhibits higher classification accuracy and generalization ability, with Precision and Dice coefficient of 93.61% and 90.76%, respectively, showing better computational efficiency and effective in segmentation tasks.

Abstract:The simple contrastive learning of sentence embedding (SimCSE) framework only uses the classification [CLS]tokens as text vectors, and it also neglects the hierarchical information within the base model, which results in insufficient extraction of semantic features from the base model output. Based on the SimCSE framework, this study proposes a method that fuses hierarchical features of pre-trained models, SimCSE with hierarchical feature fusion (SimCSE-HFF). SimCSE-HFF is based on a dual-path parallel network, using short and long paths to strengthen feature learning. The short path uses a convolutional neural network to learn local text features and perform dimensionality reduction, while the long path uses a bidirectional gated recurrent neural network to learn deep semantic information. Additionally, in the long path, an autoencoder is used to fuse features from other layers within the base model, solving the problem of insufficient extraction of output features by the model. On the Chinese and English datasets of spring tools suite-bundle (STS-B), the SimCSE-HFF method outperforms traditional methods in terms of semantic similarity Spearman and Pearson correlation metrics, showing improvements on different pre-trained models. Additionally, it also outperforms the SimCSE framework in downstream task retrieval-based question answering, demonstrating better versatility.

Abstract:Accurate prediction of wind turbine metrics is important for accurate control of turbines and the regulation of grid supply and demand. The task of forecasting these indicators can be abstracted as a task of wind power time series forecasting. Currently, deep learning models are mainly used in time series prediction models, but the strong volatility and randomness of wind power time series often prevent most models from effectively capturing the complex evolutionary characteristics of the data. To address these issues, a wind power time series forecasting method based on a progressive decomposition architecture is proposed, which first applies a neural network pooling decomposition method to simplify complex dependencies and then applies an attention mechanism to learn long-term trends. Subsequently, a multivariate fusion capture module is employed to enhance the overall multivariate correlation mining ability of the network, and it fuses the trend term and the period term to make accurate forecasts of the wind power time series. Finally, the trend and period terms are fused to make accurate forecasts for wind power time series. Experimental results show that this method can achieve up to a 24% reduction in mean squared error (MSE) for wind power time series forecasting compared to baseline models. It also exhibits stable improvements in predictive performance across multiple forecasting lengths while significantly outperforming similar models in computational efficiency.

Abstract:In the monocular image depth estimation method based on deep learning, the depth information of the image is lost during the subsampling process of the convolutional neural networks, which leads to poor depth estimation of object edges. To solve this problem, this study presents a multi-scale feature fusion method, and an adaptive fusion strategy is adopted to dynamically adjust the fusion ratio of feature maps of different scales according to feature data to make full use of multi-scale feature information. In the monocular depth estimation task using atrous spatial pyramid pooling (ASPP), the pixel information loss affects the prediction results of small objects. When using ASPP on deep feature maps, the depth estimation result is improved by fusing rich feature information of shallow feature maps. The experimental results on the NYU-DepthV2 indoor dataset show that the method proposed in this study has a more accurate prediction of object edges and significantly improves the prediction of small objects. The root mean square error (RMSE) reaches 0.389 and the accuracy (δ<1.25) reaches 0.897, which verifies the effectiveness of the method.

Abstract:In addressing issues such as feature redundancy in traditional U-shaped networks and the complexity of retinal vascular morphology, as well as challenges in segmenting fine blood vessels, this study proposes a multi-flow retinal vascular segmentation algorithm based on improved U-Net. The algorithm incorporates two feature flows, a global segmentation flow and a boundary-specialized flow. To reduce feature redundancy, the global segmentation flow replaces the traditional U-Net convolution block with a fast extraction module based on partial convolution and constructs an improved U-Net model that can efficiently extract vascular features and accelerate algorithm inference speed. To minimize noise interference and enhance the segmentation accuracy of fine blood vessels, the boundary-specialized flow utilizes morphologically generated boundary annotations as guidance. Multiple boundary extraction modules, in combination with the high-level semantic features from the global segmentation flow and boundary attention, are employed to more selectively extract vascular details, thereby strengthening the feature representation of fine blood vessels. The effectiveness of the algorithm is evaluated on the DRIVE and STARE datasets, yielding sensitivity values of 0.841 5 and 0.836 9, accuracy values of 0.970 1 and 0.971 8, and AUC values of 0.987 7 and 0.990 9, respectively. The overall performance surpassed that of existing algorithms.

Abstract:An enhanced YOLOv8n-based object detection algorithm, SFE-YOLO, is developed to tackle the issues of low detection precision for small targets in UAV aerial photography. Initially, a shallow feature enhancement module is embedded to integrate the shallow spatial details of input features with deep semantic information obtained from the neck section. This fusion strengthens the representation capability for small target features. Additionally, a global context block (GC-Block) is utilized to recalibrate this merged information, effectively suppressing background noise. Subsequently, the network’s adaptability to geometric changes is increased by substituting deformable convolutions for some standard convolutions in the C2F layer. Furthermore, the ASPPF module, incorporating average pooling technology, is integrated to augment the model’s expression of multi-scale features and to decrease miss rates. Finally, a novel weighted feature fusion method is designed. This method blends more intermediate features from the main network, enabling smoother transitions among different scale features and augmenting feature reusability through skip connections. The model’s performance is validated on VisDrone2019 and VOC2012 datasets, achieving mAP@0.5 values of 30.5% and 67.3%, respectively. These results mark improvements of 3.6% and 0.8% over the baseline YOLOv8n algorithm, demonstrating enhanced performance in UAV image target detection and notable generalization capabilities.

Abstract:The GCN-based collaborative filtering model achieves good performance in the recommendation field, but existing graph collaborative filtering learning methods usually do not distinguish the interaction relationship between users and items, which makes it difficult to mine the underlying intentions of user behavior. To address these issues, a decoupling graph contrastive learning recommendation model is proposed. Firstly, users and items are embedded into independent spaces to decouple their intentions. Secondly, during the graph propagation phase, potential semantic neighbors are discovered based on the intention features of users and items. The representation learning of structural and semantic neighbors is decoupled based on intent similarity, generating complete high-level representations for users and items. In the contrastive learning phase, nodes are randomly perturbed to create contrastive views, and contrastive learning tasks are constructed for both structural and semantic aspects. Finally, a multi-task strategy jointly optimizes the supervised task and the contrastive learning task. Experimental results on Yelp2018 and Amazon-Book datasets show that the proposed model outperforms the optimal baseline model NCL. It demonstrates improvements of 7.54% and 5.65% in Recall@20, and 8.57% and 6.28% in NDCG@20 on the two datasets, respectively.

Abstract:Multi-view clustering aims to learn more comprehensive and accurate consensus representations from the diversity information of different views to improve the clustering performance of the model. Currently, most multi-view clustering algorithms use the Hilbert-Schmidt independence criterion (HSIC) or adaptive weighting method to consider the diversity of each view from a global perspective, ignoring the learning of local diversity information between samples in each view. This study proposes a diversity-guided deep multi-view clustering algorithm to address the above issues. Firstly, the study proposes a soft clustering module integrating multi-head self-attention mechanism. Specifically, the multi-head self-attention mechanism is applied to learn global diversity, and the soft clustering fuzzy C-means algorithm is utilized to learn local diversity. Secondly, a soft clustering module is introduced into the structure of the depth map auto-encoder network to generate potential representations guided by diversity information. Then, the obtained latent representations of each view are weighted and fused to obtain consensus representations, and the spectral clustering algorithm is leveraged to cluster the consensus representations. Finally, comparative experiments and ablation experiments are conducted on three commonly used datasets. The experimental results show that the proposed clustering algorithm has good clustering performance, and the diversity information learning module can effectively improve the clustering performance of the algorithm.

Abstract:Artificial neural network (ANN) has made significant progress in many fields, but its high demand for computing resources and energy consumption limits its deployment and application on the hardware side. Spiking neural network (SNN) performs well on neural morphology hardware due to its low power consumption and fast inference characteristics. However, the neural dynamics and pulse propagation mechanism of SNN make its training process complex. Current research primarily focuses on image classification tasks. This study attempts to apply SNN to more complex computer vision tasks. This study is based on the YOLOv3 tiny network and proposes the spiking YOLOv3 model, which conforms to the SNN characteristics of the network model. It achieves higher accuracy in detection tasks and reduces the average inference time to about 1/4 of the original work. In addition, this study also analyzes the conversion errors generated during the ANN-SNN conversion process and optimizes the Spiking YOLOv3 model using a quantization activation function to reduce conversion errors. The optimized model reduces the average inference time to about half of the original and achieves lossless conversion on the VOC and UAV datasets in ANN-SNN, significantly improving the detection efficiency based on this model.

Abstract:Continual relation extraction aims to train models to learn new relations from evolving data streams while maintaining accurate classification of previously learned relations. However, due to the catastrophic forgetting problem of neural networks, the model’s ability to recognize old relations tends to decrease drastically after being trained on new relations. To mitigate the impact of catastrophic forgetting on model performance, this study proposes a continual relation extraction method based on contrastive learning and focal loss. First, the model is trained on a concatenated set of the original training set and its augmented samples to learn a new task. Second, from the training set, memory samples are selected and stored for each new relation. Then, instances from the activation set are contrasted with all known relation prototypes to learn the old and the new relations. Finally, memory reconsolidation is performed using the relation prototypes and focal loss is introduced to improve the model’s distinction between similar relations. Experiments are conducted on the TACRED dataset, and the results show that the method proposed can further alleviate catastrophic forgetting and improve the model’s classification ability.

Abstract:RS (Reed-Solomon) code is most widely adopted in distributed storage systems that support erasure coding. For an RS(k,m) coding stripe, a common approach to store it is to distribute one fragment to one node. Such an approach could cause network congestion when a node fails since the system needs to read fragments across multiple nodes before it can decode and rebuild the lost data. The system would be in a fragile period for a long time when a great amount of data recovery is taking place. During this period, the system would suffer from lower failure tolerance capability, lower throughput, and higher read/write latency constantly. LRCRaft is an optimized version of Raft based on local reconstruction code (LRC). By introducing LRC, dynamic log replenishment, state machine purge, and fragment version consistency to Raft, LRCRaft can reduce read/write latency and the time consumed for node failure recovery. The results of our experiments indicate that compared to Raft, LRCRaft can reduce the time for a single node recovery by up to 49.25%–74.97% in different recovery modes.

Abstract:Unsupervised feature selection based on spectral clustering mainly involves the correlation coefficient matrix and the clustering indicator matrix. In previous studies, scholars have mainly focused on the correlation coefficient matrix, designing a series of constraints and improvements for it. However, focusing solely on the correlation coefficient matrix cannot fully learn the intrinsic structure of data. Considering the group effect, this study imposes the F-norm on the clustering indicator matrix and combines it with spectral clustering to make the correlation coefficient matrix learn more accurate clustering indicator information. The two matrices are solved through an alternating iteration method. Experiments on different types of real datasets show the effectiveness of the proposed method. In addition, experiments show that the F-norm can also make the method more robust.

Abstract:The detection speed and accuracy of detecting targets ahead during vehicle operation have always been a focus of research in the field of autonomous driving. For existing object detection algorithm models, in complex traffic environments, traditional models are prone to false positives and missed detections when facing overlapping targets. Significantly improving detection accuracy can also lead to increased computational demands, resulting in slower processing speed and decreased real-time performance. This article proposes an improved algorithm based on the YOLOv5 model. Firstly, the MobileNetV3 network is adopted to replace the C3 backbone network in the original model, achieving a lightweight network while improving the model’s response speed. Secondly, a non-maximum suppression algorithm, Adaptive-EIoU-NMS, is proposed to improve the recognition accuracy of overlapping targets. Finally, the K-means++ clustering algorithm is used to replace the original clustering algorithm and generate more accurate anchor boxes. Experimental results show that the improved model achieves an average detection accuracy of 90.1% and a detection speed of 89 frames per second (f/s). The experimental results confirm that the enhanced model significantly improves both detection accuracy and speed for complex scene detection.

Abstract:Compared to traditional supply chains, the large-scale, digitalized industrial interconnected intelligent manufacturing supply and demand network has stronger response and adjustment capabilities as well as risk prevention and recovery capabilities. However, it also faces a greater variety of risks and has broader risk transmission pathways, more easily threatening its robustness. Accurately describing the dynamic propagation process of fault risks within the network is fundamental to enhancing its robustness. Firstly, a model of the industrial interconnected intelligent manufacturing supply and demand network with multiple industrial communities is constructed. Secondly, by considering the relative correlation between business nodes, a risk propagation model with relative fault probabilities is built. Then a fault recovery model that considers both the recovery probability and recovery period is established based on the importance of nodes. Finally, a network is constructed based on an improved gravity model, and the relative connectivity rate R of the network is used as an indicator to simulate and analyze cascading failures under different fault and recovery scenarios. The simulation results indicate that there are critical values in all four sets of different fault and recovery scenarios that lead to R being in an unstable state in the long run. The parameters η and μ have certain marginal effects on the value of R. When the network’s fault propagation capability is fixed, the weaker the recovery capability, the more pronounced the oscillation of R, and the larger the scale of network impact. Conversely, with a certain recovery capability, the stronger the fault intensity, the more pronounced the oscillation of R, and the larger the scale of network impact.

Abstract:The emergence of network function virtualization (NFV) technology allows network functions are provided by virtual network functions (VNFs) to improve network flexibility, scalability and cost-effectiveness. However, an important challenge for NFV is how to efficiently place VNFs in different network locations and chain them to steer traffic while minimizing energy consumption. In addition, in the face of network quality of service requirements, improving the service acceptance rate is also critical to network performance. To address these issues, in this study we investigate VNF placement and chaining (VNFPC) in NFV to maximize the service acceptance rate while optimizing the energy consumption trade-off. Therefore, an energy-efficient VNFPC method based on Actor-Critic deep reinforcement learning (DRL), called ACDRL-VNFPC, is designed in NFV. The approach applies adaptive sharing scheme to achieve energy savings by sharing the same type of VNFs among multiple services and sharing the same server among multiple VNFs. The experiment results show that the proposed algorithm effectively trades off the energy consumption and service acceptance rate, and the execution time is also optimized. Compared with the baseline algorithm, ACDRL-VNFPC improves the performance in terms of service acceptance rate, energy consumption and execution time by 2.39%, 14.93% and 16.16%, respectively.

Abstract:In previous machine reading comprehension models, there were some problems, such as single-text feature extraction and incomplete interactive information between text and questions, which led to insufficient text understanding. This study proposes a machine reading understanding model with multi-level information fusion, which can obtain text information at multiple levels by using different methods in different locations. The model uses the dilated convolutional network to capture the global information of the text. Bi-directional attention mechanism and self-attention mechanism are used to fuse the interactive information between text and questions. Finally, the answer and its corresponding supporting sentence are predicted through the pointer network. The joint F1 values of the model trained on the CAIL2019 and CAIL2020 reading comprehension datasets reach 50.09% and 58.44% respectively, which achieves significant performance improvement compared with other baseline models.

Abstract:As an important part of China’s space transportation system, the solid launch vehicle has characteristics of storage and transportation of the whole rocket and rapid launch response, and is widely favored in the field of military, civilian, and commercial launch. In response to the alignment error problem caused by actuator deformation of vehicle-mounted lifting equipment in the automated transfer of launch vehicles, this study refines an automatic transfer method based on a two-stage alignment model. This method aims to solve the closed-loop detection deficiency of alignment error caused by mechanism deformation that conventional transfer methods cannot effectively handle. The proposed method is verified by using Monte Carlo simulation. The results show that compared with conventional transfer methods, the proposed method has a success rate of about 96% for a single transfer, which solves the accuracy problem of automatic transfer alignment under heavy loads and large deformation. The transfer accuracy is good and can ensure accurate docking between rockets and vehicles.