Abstract:Aiming at the lack of global semantic consistency constraints and insufficient acquisition of local features of the current deep learning algorithms in the process of image restoration of broken murals， resulting in the restored murals being prone to boundary effects and blurring of details， this paper proposes a bidirectional autoregressive Transformer with fast Fourier convolutional enhancement of murals restoration method. First， a global semantic feature repair module based on the Transformer structure is designed， and an improved multi-head attention global semantic mural repair module is proposed using the bidirectional autoregressive mechanism with masked language modeling （MLM） to improve the repair capability of global semantic features. Then， a global semantic enhancement module consisting of gated convolution and a residual module is constructed to enhance the global semantic consistency constraint. Finally， the local detail repair module is designed， which adopts large kernel attention （LKA） and fast Fourier convolution （FFC） to improve the ability of capturing detailed features while reducing the loss of local detail information， so as to enhance the consistency of the local and overall features of the repaired murals. The experimental results of the digital restoration of real Dunhuang murals show that the proposed algorithm can effectively restore the structure and texture of the murals， and the subjective visual effect and objective evaluation indexes are better than the comparative algorithms.

Abstract:The current CNN-based methods exhibit satisfactory performance in fog removal， but their network robustness is compromised due to the intricate haze distribution and challenging dataset collection. Consequently， there is a significant loss of texture details during the fog removal process and severe overfitting issues on small-scale data sets. To address these challenges， we propose a two-branch structure incorporating space-frequency joint techniques. The upper branch focuses on capturing finer texture details by utilizing three-level wavelet transform to extract features in the frequency domain. Meanwhile， the lower branch enhances network generalization by employing domain migration method to incorporate additional prior information from airspace and leveraging Res2Net as its core component. Finally， the proposed model is trained on the NH-HAZE dataset and evaluated for generalization ability using the I-HAZE and NTIRE 2023 datasets. Furthermore， to ensure fairness in comparison experiments， all competing algorithms are also trained using the NH-HAZE dataset. Experimental results demonstrate that our proposed network significantly improves both detail texture recovery capability and generalization performance.

Abstract:In view of the significant texture difference between fake faces and real faces， this paper proposes a face forgery detection model based on texture features. Firstly， ResNet18 is used as the backbone network， and combined with the channel attention mechanism and residual network to solve the problem of network degradation， in order to establish the connection between channels to extract deep features. Secondly， the autocorrelation matrix is used to quantify the correlation between image blocks， and the features of different scales in the image are captured to obtain global statistical features. Finally， the Gabor filter is introduced after each pooling layer of the autocorrelation module to extract the local texture features of the image， providing a comprehensive description of the image content， and the Softmax function is used to perform hierarchical classification. Experimental results show that this method effectively improves the detection accuracy for fake images edited by different image enhancement methods.

Abstract:The core of point cloud registration is to estimate the transformation matrix. There is partial overlap， high noise， and density difference between two point cloud pairs. The existing methods cannot accurately solve the problem of feature alignment between significant point cloud correspondences. Therefore， a significance peak and feature alignment network （SPFANet） is proposed to achieve point cloud registration from coarse to fine one. SPFANet consists of three parts： multi-significance peak detector， coarse registration， and fine registration. Firstly， the multi-significance peak detector introduces a re-weighted peak loss method based on descriptor variance and overlap score to remove non discriminatory and non overlapping key point clouds. Secondly， the coarse registration stage detects the complementary key point sets to compute the coarse registration scheme. Finally， the fine registration stage introduces a feature metric framework with a forward-backward transform to refine the coarse registration scheme and achieve efficient point cloud registration. The effectiveness of SPFANet is validated through experiments on the same source 3DMatch dataset and cross-source 3DCSR dataset.

Abstract:To solve the problems of insufficient understanding and fuzzy feature boundary segmentation of multi-target categories small-scale feature semantic information in aerial remote sensing images in complex background， this paper designs a segmentation model that integrates the features of the backbone network information and classifies and reconstructs the features to improve the segmentation effect. The model takes Swin-Transformer as the coding structure and utilizes its ability to understand global semantic information for feature extraction. The segmentation of small-scale target features is refined by the designed information grouping reconstruction convolution （IGRM） and channel classification reconstruction convolution （CRRM）， which classify and reconstruct the extracted features by the amount of information. Finally， by integrating the up-sampling and down-sampling connections， the reconstructed features are fused with the features extracted by the encoder to form a multi-scale feature aggregation block to output the segmentation results. The refined reconstruction of small-scale target features is realized in multi-target scenarios with complex backgrounds， and high-quality segmentation maps are generated to improve the segmentation accuracy. Experimental results on the ISPRS Potsdam and ISPRS Vaihingen datasets show that the average intersection and merger ratio （mIoU） is 87.15% and 82.93%， respectively， and the overall accuracy （OA） is 91.53% and 91.4%， respectively. To verify the generalization ability of the model for small-scale target feature extraction in multi-target categories， this paper also designs a comparative experiment for the category of carts in complex backgrounds. The experimental results show that the mIoU on the UAVid dataset reaches 67.86%.

Abstract:The existing object detection algorithms， influenced by complex environments and the complexity of network models， face challenges in effectively detecting distant unmanned aerial vehicles （UAVs）. This paper proposes an improved unmanned aerial vehicle （UAV） target detection algorithm based on YOLOv8. First， to address the challenge of detecting small unmanned aerial vehicle targets at long distances， a new ultra-small object detection layer is proposed， which integrates shallow features. In this approach， the largest target detection layer is removed to optimize target scale focus and reduce network complexity. Second， the Backbone part incorporates the GhostConv module to further decrease the model’s parameter count. Then， in the Neck part， the LSKA attention mechanism is integrated by replacing the Bottleneck section in the C2f module with LSKA， designing a new C2f-LSKA module to replace some C2f modules in the Neck， enhancing the model’s contextual awareness and spatial information processing ability. Lastly， WIoUv3 is used as the boundary loss function to further improve model accuracy. Experimental results show that， compared with the original model， the improved model increases precision （P） by 5.0%， recall （R） by 11.9%， and mAP@0.5 by 9.5% on a custom UAV dataset， while reduces the model’s parameter count and size by 68.9% and 65.1%， respectively.

Abstract:To solve the problems of the existing target tracking algorithms， such as inability to extract deep-level features， failure to fully exploit cross-modal information， and weak representation of target features， a feature fusion shift Siamese network for RGB-T target tracking is proposed. First， a target tracking framework based on the visible modal SiameseRPN++ is designed to extend the infrared modal branch， in order to obtain a multimodal target tracking framework. Moreover， the improved ResNet50 network with adjusted stride as a feature extraction network enables the acquisition of deep-level features of the target. Subsequently， a multimodal feature interactive learning module （FIM） is designed to leverage the discriminative information from one modality to guide the learning process of target appearance features in the other modality. By mining the cross-modal information within the feature space and channels， the module enhances the network’s attention towards foreground information. Thereafter， a multimode feature fusion module （FAM） is designed， which calculates the degree of feature fusion between the input visible light image and the infrared image， enabling spatial fusion of significant features from different modalities to effectively eliminate redundant information and reconstructing multimodal images by employing a cascade fusion strategy. Finally， a feature space shift module （FSM） is designed， which divides the feature maps of the infrared modal branches and shifts them in four different directions to enhance the edge representation of the heat source target. Extensive experiments on two RGB-T datasets thoroughly validate the effectiveness of the proposed algorithm， while ablation experiments demonstrate the superiority of each designed module.

Abstract:To address the challenges of low offloading success rates and inefficient data transmission in the internet of vehicles （IoV）， this paper proposes a multi-layer distributed dynamic offloading strategy for edge computing tasks in IoV based on multi-agent deep reinforcement learning. Firstly， a multi-layer distributed internet of vehicles edge computing system model is designed by integrating software defined network and mobile edge computing. The system model can realize collaborative scheduling optimization at different levels， which can better meet the needs of dynamic allocation of mobile vehicle resources and real-time processing of tasks. Then， considering the success rate of offloading and data transmission rate of vehicle computing tasks， a multi-agent deep reinforcement learning algorithm framework is proposed. The algorithm framework uses collaborative learning of multi-agent systems to enable the vehicle edge system to independently select the optimal task offloading decision. At the same time， the optimization of the action space search and the priority experience replay mechanism were introduced to further improve the effective search of the action space and the stability and accuracy of the task offloading decision. Finally， based on the above algorithm framework and optimization mechanism， a multi-layer distributed vehicle task offloading decision optimization algorithm is proposed. The algorithm can ensure that the vehicle can complete the computing task offloading with the minimum task transmission time and effective offloading success rate according to the current network status and task size. Simulation results show that， compared with the existing offloading methods， the proposed method improves the success rate of computing task offloading by 5%~20% and the efficiency of data transmission by 17.8% on average.

Abstract:Unmanned aerial vehicle （UAV） communication faces challenges such as path loss and intergroup interference. To meet the discrete users’ communication needs， achieve static deployment of UAV networks， and maximize energy efficiency， this paper studies a multi-ratio concave-convex fractional programming problem. A convex optimization cooperative swarm intelligence strategy is proposed， which decouples the original problem into separate power control and height optimization problems， solving them iteratively. Firstly， a line-of-sight （LoS） probability average path loss model is introduced to study the relationship between deployment height and horizontal distance， as well as the three-dimensional deployment problem through pitch angles. Secondly， a quadratic transformation is utilized to decouple the original problem， aiming to enhance system energy efficiency under the LoS probability link. Finally， a fast feedback particle swarm algorithm is proposed for accurate deployment of heights， addressing the complex multi-objective cooperative optimization problem. Simulation results demonstrate that， under the proposed model， the strategy achieves the balance between algorithm complexity and accuracy， enabling efficient and accurate deployment of UAV base stations.

Abstract:The existing deep learning-based source code vulnerability detection methods mainly focus on the feature learning of a single programming language， and it is difficult to effectively detect the vulnerabilities caused by the association and invocation of code units in software projects of hybrid programming languages. To address this issue， a deep learning-based hybrid language vulnerability detection method DL-HLVD is proposed. Firstly， the BERT layer is used to convert the code text into low-dimensional vectors， which are then used as inputs to the bidirectional gated loop unit to capture the contextual features， and the conditional random field is used to capture the dependency between adjacent labels. Secondly， functions from different types of programming languages are identified as named entity recognition in the hybrid software and reconstructed with the program slicing results to reduce the loss of syntactic and semantic information in the code characterization process. Finally， the bidirectional long short-term memory network model is designed to extract the vulnerability code features and realize the vulnerability detection of hybrid language software. The comprehensive experimental results on the SARD and CrossVul datasets show that the comprehensive recall rate of DL-HLVD on the two types of vulnerability datasets is 95.0%， and the F1 value reaches 93.6%， which is improved in all indicators compared with the VulDeePecker， SySeVR， and Project Achilles. It demonstrates that the DL-HLVD method can improve the comprehensive performance of source code vulnerability detection in hybrid language scenarios.

Abstract:To improve the naturalness and robustness of synthesized prosody， a autoregressive speech synthesis model based on phoneme-level prosody modeling is proposed. This model enhances prosody modeling from two aspects： inter-word pauses and phoneme durations. To enhance the diversity and accuracy of inter-word pauses， a pause prediction module is proposed at the text frontend. This module predicts multiple pause labels based on the original text， thereby providing accurate references for pause duration modeling in speech synthesis. To enhance the naturalness of phoneme durations， a duration prediction module is proposed. This module predicts a mixture Gaussian distribution for each phoneme and obtains diversified phoneme durations through random sampling. To stabilize phoneme duration modeling in the autoregressive model， an attention-based discrimination module is proposed. This module is applied at each time step of the autoregressive process and avoids alignment disorder through attention and discrimination mechanisms. Experimental results demonstrate that the three proposed modules effectively enhance the naturalness and robustness of prosody modeling， thereby improving the quality of speech synthesis.

Abstract:Due to the characteristics of electronic medical records （EMRs）， such as the diversity of data types and temporal irregularity inherent， most existing deep learning-based methods cannot simultaneously capture static correlations between different types of clinical data and dynamic temporal dependencies between visits during the feature learning process. To address this issue， this paper proposes a disease prediction model based on multi-domain graph neural network. In this model， a temporal feature learning module that combines code level attention and time aware LSTM is first utilized to obtain the initial feature representation of patient visits. Then， based on the correlation and time interval information between different visits， a visit affinity graph and a visit sequence graph are constructed， and a graph convolutional neural network is used to mine the static and dynamic semantic associations between visit records from these graphs. Finally， a multi-domain feature fusion module based on self-attention mechanism is utilized to combine temporal features and semantic association features to obtain the final patient fusion representation for future disease prediction. The experimental results on two real clinical datasets show that our method outperforms other existing methods and achieves higher prediction accuracy.

Abstract:To improve the accuracy of steel defect detection， a defect detection algorithm YOLOv5s-FNCE based on YOLOv5s is proposed. Firstly， a novel NAMAttention attention mechanism is added to the backbone feature extraction network to improve the perception and differentiation of the target； and a new C3-Faster is proposed to extract the features； the positional convolutional CoordConvs is introduced in the feature fusion network and at the output to enhance the semantic perception ability and global perception ability of the target； and finally， a new loss function Focal-EIoU is introduced to accelerate the convergence speed and improve the regression accuracy. Experimental results show that the mean average accuracy of the YOLOv5s-FNCE algorithm on the steel surface defects dataset reaches 75.1%， which is 1.7% higher than that of the original YOLOv5s， the detection speed is increased by 20.5%， which proves that the algorithm can effectively improve the detection speed and accuracy in steel defect detection.

Abstract:In this paper， a landslide displacement prediction model integrating extreme gradient boosting and optimized support vector regression is proposed by using extreme gradient boosting and support vector regression， and combining the advantages of hunter-prey optimization algorithm. Firstly， extreme gradient boosting （XGBoost） is used for the preliminary prediction of landslide displacement， and then hunter-prey optimizer （HPO） is used to optimize support vector regression （SVR）. A combined prediction model （HPO-SVR） is constructed by optimizing the hyperparameters of SVR using HPO to correct the prediction results of XGBoost. The validation of two sets of landslide displacement measured data shows that the HPO algorithm obtains a more reasonable hyperparameter of SVR through the dynamic optimization strategy of constantly updating the positions of the hunter and the prey. Relative to the combined prediction models of XGBoost， SVR， and its combination with particle swarm optimization algorithm， genetic algorithm， and HPO， the combined XGBoost-HPO-SVR model achieves good results in predicting the displacements of Yangwashan landslide and Tuojiashan landslide， with mean square errors of 3.505 and 0.550， and mean absolute errors of 1.357 and 0.538， respectively.

Abstract:This study focuses on the governance problem of e-commerce platform big data killing under the role of government regulation and consumer rights defense， introduces the rights defense cost sharing mechanism and punitive damages， constructs a tripartite evolutionary game model of government-e-commerce platform-consumers， and explores the factors of effective inhibition of big data killing on e-commerce platforms. The study shows that 1） e-commerce platforms will not give up big data killing； 2） improving the success rate of regulation， increasing government punishment， and improving the coefficient of punitive compensation， the rate of rights defense and the success rate of rights defense can effectively inhibit big data killing； 3） sharing the cost of rights defense cannot effectively inhibit e-commerce platforms big data killing.

Abstract:The existing EDA tools address the global placement problem of very large scale integration （VLSI） physical design by minimizing the sum of half-perimeter wirelength （HPWL） under density constraints. However， the non-differentiability of HPWL renders gradient-based advanced optimization methods inapplicable directly to global placement. Consequently， the weighted-average wirelength （WAWL） model is often employed to approximate HPWL， but it struggles to achieve a balance between smoothness and accuracy. This paper introduces an improved self-adaptive weighted-average wirelength （SaWAWL） model. It dynamically adjusts the weighted factors γ for each wire’s actual length， ensuring both smoothness and reduced error in fitting HPWL. The proposed model enhances the quality of global placement. A global placer based on this model is implemented and validated on the DAC 2012 open benchmark. The results indicate a 3.69% reduction in the total sum of half-perimeter wirelength.

Abstract:Nano-SIC particle/whisker hybrid reinforced 8009Al/6061Al composites （AMC-DNC） with dual network structure were successfully prepared by spray granulation and discharge plasma sintering. At the same time， the SiC reinforced phase is uniformly distributed in 8009Al/6061Al matrix （AMC-US）. In this study， the microstructure and friction and wear properties of the two composites were investigated. The friction coefficient and wear rate of AMC-DNC and AMC-US under different load and rotation speed were measured. The wear morphology was observed by scanning electron microscope， and the influence of dual network structure on the wear mechanism of composite was analyzed. The experimental results show that the density of AMC-DNC and AMC-US prepared by discharge plasma sintering method is 98.1% and 99.2%， respectively. Under low rotation load and rotation low speed conditions， the main wear mechanism of AMC-DNC is abrasive wear， in which the separation of particles leads to the rupture of oxide film， resulting in the surface damage of the material. However， under higher load and rotation speed conditions， the wear mechanism changes to peeling wear. Under the same friction and wear experimental conditions， AMC-DNC shows better friction and wear performance than AMC-US， indicating that the design of dual network stucture plays a significant role in improving the wear resistance of composite materials.

Abstract:The effects of dual aging process on microstructure and mechanical properties of as-cast Al-Si-Mg-Cu alloy were investigated by scanning electron microscopy， transmission electron microscopy and tensile test at room temperature. The dual aging process used in this study was low temperature pre-aging and high temperature final aging. The results showed that more precipitation nucleation sites were formed in the low temperature pre-aging stage of the alloys. After high temperature final aging， more θ" phases and some β" phases were precipitated in the alloys. When the pre-aging was 155 °C/8 h， the size of θ" phase gradually increased with the increase of final aging temperatures， while the morphology of the β" phase did not change significantly. The tensile strength of the alloys increased first and then decreased with the increase of the final aging temperatures. The elongation of the alloys decreased significantly with the increase of the final aging temperature， and 165 °C was the better final aging temperature of the alloy. When the dual aging process was 155 °C/x h+165 °C/2 h， the number of θ" phase and β" phase in the alloy gradually increased with the increase of pre-aging time. When the dual aging process was 155 °C/8 h+165 °C/x h， the size of θ" phase increased gradually with the increase of the final aging time. The number and size of β" phase in the alloy did not change significantly before the final aging time was 8 h. Continuing to extende the final aging time， the β" phase also began to grow. The dual precipitated phase strengthening of θ" phase and β" phase can be realized by optimizing the dual aging process. The tensile strength of the alloy can reach 380 MPa after 155 °C/8 h+165 °C/2 h dual aging， while the elongation of the alloy was increased to 3.4 % after 155 °C/8 h+ 165 °C/8 h dual aging.

Abstract:Taking the super heat-resistant aviation gear steel 12Co14Ni6Cr5Mo4VNb cylindrical gear as the research object， the high-frequency pulse method was used to conduct bending fatigue performance tests on gears with different surface modifications. Through experimental data analysis and gear tooth failure mechanism analysis， the influence of different surface modification layers on the load-bearing performance of super heat-resistant aviation gear steel gears was revealed. The research results show that ordinary shot peening has no significant strengthening effect on the surface of super heat-resistant aviation gear steel gears. Super shot peening can significantly increase the residual compressive stress on the surface of super heat-resistant aviation gear steel gears. Compared with ordinary shot peening gears， the bending fatigue limit of super shot peening gears is increased by 36.04%； Super heat-resistant aviation gear steel gears are highly sensitive to the tooth surface roughness， gears with roughness Ra≤0.2 showing a 16.58% increase in bending fatigue limit compared with gears with a roughness of 0.3≤ Ra≤0.4. In the bending fatigue test， the microstructure of carburized layer near the surface of the gear is high-carbon martensite and finely dispersed granular carbides， and it gradually transits to ultra-low carbon flat noodles martensite with the increase of carburized layer depth. The failure mode of the test gears is fatigue fracture of the gear teeth， with clear fatigue bands visible on the fracture surface. The fatigue cracks of the super shot peening gears all originate from the sub surface of the tooth root fillet， while the fatigue cracks of ordinary shot peening gears mostly originate from the surface of the tooth root fillet.

Abstract:The aim of this study is to develop an efficient heat storage and temperature regulating material suitable for a variety of thermal energy storage systems. The capric acid-lauric acid （CA-LA） low eutectic mixture was encapsulated in a poly （ethyl acrylate） （PEA） shell through emulsion polymerization to form PEA/ （CA-LA） phase change material microcapsule （MEPCM）. Analysis results demonstrate that the prepared MEPCM exhibits a uniform and smooth spherical appearance， with no chemical reaction occurring between the core and shell materials. At a shell-core ratio of 1∶1.5， the latent heat of melting and solidification are 81.85 J/g and 88.68 J/g， respectively. The latent heat storage capacity of MEPCM increases with an increase in core material ratio. Through thermogravimetric analysis and leakage testing， it is observed that MEPCM maintains excellent latent heat storage and release ability below 160 ℃. Even after 200 thermal cycles， it still demonstrates good thermal stability and practical application value.