Abstract:To solve the problem that the existing deep learning algorithms do not fully consider the consistency of the information between the damaged area and the intact area when repairing mural images， which leads to boundary effects and texture blur in the repair results， we proposed a progressive mural inpainting algorithm combining feature reasoning and semantic enhancement. Firstly， the progressive structure of the region was designed to realize the progressive contraction of the region to be repaired. Then， the feature reasoning module was used to iteratively fill the feature values of the missing pixels， reduce the reconstruction error of the mural restoration， and enhance the correlation between the damaged area and the intact area of the mural. Finally， the feature maps of each layer were adaptively fused， and the semantic enhancement module was used to transfer the texture details， so as to improve the consistency of the mural completion area and the whole. The digital restoration experiments of Dunhuang murals show that the restored murals by the proposed method have better consistency of texture details， and are superior to the comparison algorithms in subjective and objective evaluation indicators.

Abstract:The current defogging algorithm cannot solve the problem of uneven haze image defogging， so this paper proposes a wide self-attention fusion conditional generation against network image defogging algorithm. The wide self-attention mechanism is added to the algorithm， so that the algorithm can automatically assign different weights to the features of different haze regions. The feature extraction part of the algorithm adopts the DenseNet fusion self-attention network architecture. Under the premise of ensuring the maximum information transmission between the middle layers of the network， the DenseNet network directly connects all layers to obtain more context information and make more effective use of the extracted features. Fusion of self-attention can learn complex nonlinearity from the features extracted from the encoder part， and improve the ability of the network to accurately estimate different haze. The algorithm uses Patch discriminator to enhance local and global consistency of defogging images. The experimental results show that the qualitative comparison of the algorithm network on NTIRE 2020， NTIRE 2021 and O-Haze datasets has better visual effects than other advanced algorithms. In the quantitative comparison， compared with the best performance of the selected advanced algorithms， the peak signal-to-noise ratio and the structural similarity index increases by 0.4 and 0.02， respectively.

Abstract:Affected by insufficient nighttime illumination， some content in night vision imaging is prone to missing or blurring， resulting in poor colorization. To address this issue， this paper proposes a colorization algorithm of night vision images based on generative adversarial network， where the image colorization in the blurred area is improved through texture detail prediction. Firstly， in the blurred area restoration， down-sampling is used to gradually reduce the proportion of the blurred image patches. What’s more， gradient adjustment predictor is used to predict the pixel values around the blurred image patches so as to continuously enhance and remedy the blurred texture details. Then， in the colorization process， we use the super-resolution imaging and the advanced adversarial network colouring model to obtain a clearer color image through minimizing the brightness and texture distortions. Experimental results show that， the PSNR of gray image increases by 0.33 dB on average after the distortion and enhancement in the blurred area. Compared with the previous advanced colorization methods， the proposed method can give the grayscale night vision image richer and more natural colors， and express the details of the image more clearly. It helps to improve the efficiency of target detection and recognition.

Abstract:Face inpainting aims to repair the missing regions in the input face and generate satisfactory high-quality results. However， it is difficult to directly repair the incomplete face when the missing region is large， and the global context awareness ability of the inpainting network determines the quality of the inpainting results. Therefore， a dual self-attention mechanism that combines soft attention and hard attention is proposed to improve the global context awareness of the inpainting network. This module obtains soft and hard attention features by calculating the global similarity and can adaptively fuse the attention features. Besides， a multi-scale generative adversarial network is proposed to promote the inpainting network to generate more high-quality inpainting results， by strengthening the supervision of inpainting results. Experimental results demonstrate that our method is superior to five state-of-the-art comparison methods from both quantitative and qualitative experiments.

Abstract:To solve of the defects of the existing image inpainting algorithms， such as the lack of contextual information and effective perceptual field， leading to poor performance when recovering large random damages and being restricted to low-resolution images， a parallel fast Fourier convolution generation inpainting algorithm based on residual transformer is proposed. Firstly， a transformer-based improved residual network module is proposed to extract the texture features from the image to be inpainted. Subsequently， a parallel fast Fourier convolution module is designed to enhance highly effective sensory field and capture the structural information from the corrupt areas. Finally， the gated dual-feature fusion module is developed to exchange and combine the structural and texture components of the images to fuse the contextual features and improve the fine-grained nature of the generated textures. Qualitative and quantitative experiments are conducted on two public datasets， and the experimental results show that the proposed algorithm can effectively restore random irregular large broken regions with complex structures and fine textures， generate high-fidelity images with reasonable structures， fine textures and rich semantics， and can be used for target removal of high-resolution images.

Abstract:Aiming at the problems of insufficient feature extraction， inefficient utilization， and insufficient ability to reconstruct high-frequency details in the process of image high-frequency information restoration， an image super-resolution reconstruction algorithm based on information distillation cascade telescopic network is proposed in this paper. Firstly， a feature scalable information distillation block is constructed， which solves the problem of insufficient feature extraction in the nonlinear mapping process of image super-resolution reconstruction of information distillation by expanding the feature receptive field of input information and using channel attention to extract interested information. Then， a cascaded residual superposition mapping block is designed， which combines multiple residual blocks. By leading out the residual part in the residual structure and cascading superposition， the transmission of information between information distillation blocks is increased， so that the extracted feature information contains more details. Experimental results show that the reconstructed image of this algorithm is clearer than that of other comparison algorithms， and the peak signal-to-noise ratio （PSNR） and structural similarity （SSIM） are greatly improved．

Abstract:The feature extraction process in deep supervised Hash image retrieval has been dominated by the convolutional neural network architecture. However， with the application of Transformer in the field of vision， it becomes possible to replace the convolutional neural network architecture with Transformer. In order to address the limitations of existing Transformer-based hashing methods， such as the inability to generate hierarchical representations and high computational complexity， a deep supervised hash image retrieval method based on Swin Transformer is proposed. The proposed method utilizes the Swin Transformer network model， and incorporates a hash layer at the end of the network to generate hash encode for images. By introducing the concepts of locality and hierarchy into the model， the method effectively solve the above problems. Compared with 13 existing state-of-the-art methods， the method proposed in this paper has greatly improved the performance of hash retrieval. Experiments are carried out on two commonly used retrieval datasets， namely CIFAR-10 and NUS-WIDE. The experimental results show that the proposed method achieves the highest mean average precision （mAP） of 98.4% on the CIFAR-10 dataset. This represents an average increase of 7.1% compared with the TransHash method and an average increase of 0.57% compared with the VTS16-CSQ method. On the NUS-WIDE dataset， the proposed method achieves the highest mAP of 93.6%. This corresponds to an average improvement of 18.61% compared with the TransHash method， and an average increase of 8.6% in retrieval accuracy compared with the VTS16-CSQ method.

Abstract:In order to solve the problems of low detection accuracy， slow training convergence speed of existing adversarial example detection methods， a method of adversarial example detection based on image denoising technology and image generation technology is proposed. The detection method converts the adversarial example detection problem into an image classification problem. It does not need to know the structure and parameters of the attacked model in advance， and only uses the semantic information and classification label information of the image to determine whether the image is an adversarial example. Firstly， a shifted masked auto-encoder based on swin-transformer and vision-transformer is used to remove the adversarial noise in the image and restore the semantic information of the image. Then， the image generation part based on conditional generative adversarial networks with gradient penalty is used to generate images based on image classification label information. Finally， the output of the images in the first two stages is input into the convolutional neural network for classification. By comparing the classification results of the denoised images and the generated images， it is determined whether the detected images are adversarial examples. The experimental results on MNIST， GTSRB， and CIAFAR-10 datasets show that the proposed adversarial example detection method outperforms the traditional detection methods. The average detection accuracy of this method is improved by 6%~36%， the F1 score is increased by 6%~37%， and the training convergence time is reduced by 27%~83%， respectively.

Abstract:Aiming at the difficulty of real-time detection and limited computing resources due to the scale change of drones in public safety areas such as playgrounds and parks， a network dynamic real-time detection method for drones， YOLO-Ads， is proposed to increase the robustness of network ability to detect drone change. Firstly， the drone data set was built independently. Secondly， a new MDDRDNet network was established with the lightweight network as the backbone to reduce the complexity of model calculation， and the coordinated attention mechanism module was introduced to strengthen the network’s attention to space and channels. Then， the mean clustering algorithm is used to regenerate the prior frame， and the optimization method combining multiple probes and multiple data sets is used in the selection of the prior frame， so that the regenerated prior frame matches the drone better. The idea of feature fusion and residual error establishes a new detector head to adapt to the detection of smaller-scale drones. Finally， a class activation mapping module is introduced into the detection module to generate a heat map， so as to observe the sensitivity of the network to changes in the scale of drones. At the same time， comparative experiments are conducted with the current mainstream networks SSD， CenterNet， YOLOv5， YOLOx， etc.， and different backbone networks ResNet， EfficientNet， VGGNet， etc. The experimental results show that the newly proposed algorithm has an average accuracy of 96.62% in the detection of scale-changing drones. Compared with the YOLOv4 algorithm， it is increased by 1.88% . The detection speed is 47 frames per second， which is 19 frames higher than that of the YOLOv4 algorithm. The memory occupied by the model is about 10.844 M， which is about one-sixth of the original memory. It reflects the timeliness and robustness of the method.

Abstract:With the continuous development of image editing technologies， it is particularly significant to develop image forensics technologies for image content security. Most existing forensics methods concentrated on single image manipulation detection but with weak robustness and no considerations on tampering location. This paper presents a multi-manipulation image forgery detection method based on convolutional neural network. In this network， a convolution flow based on residual block is constructed to extract manipulation features. Then， a multi-scale feature fusion module is designed to achieve operational feature fusion at different scales. Finally， the fused manipulation features are fed into the multi-branch prediction module， predicting the type and location of each utilized manipulation as the multi-manipulation detection results. An image dataset produced by multiple typical image manipulations is built to train and test the proposed network. The experimental results show that the proposed scheme can recognize the type of tampered manipulations and locate the tampered area more accurately with fewer parameters， and has better robustness to common image post-processing operations， compared with the state-of-the-art object detection networks.

Abstract:Objective factors lead to poor contrast and blurred edges of the lesion area in eye optical coherence tomography （OCT） images obtained clinically. To address these issues， a super-resolution reconstruction method for single eye OCT image， named EOTRN， is proposed by referring to the information processing mechanism of the eagle vision system. It imitates the idea of gradually expanding the receptive field of the eagle optic tectum， and excavates advanced semantic features from both the vertical and horizontal dimensions step by step. In the vertical dimension， EOTRN utilizes dilated convolutions， dense connection and a channel attention mechanism to gradually expand the receptive field. This process propagates the characteristics of different network layers， enabling competition or cooperation among different channel features. As a result， the preliminary extraction of advanced semantic features of low-frequency signals is achieved. In the horizontal dimension， EOTRN eliminates redundant information from the advanced semantic features， corrects and highlights salient information， and enhances the texture and contour features of the lesion areas using 64 characteristic subspaces. Finally， the underlying semantic features and the advanced semantic features are upsampled and deeply reconstructed to obtain high-definition images. the Experiments show that， for the third test set with scale factor ×4， compared with EMASRN， EOTRN achieves 0.96% increase in PSNR and 1.36% increase in SSIM values. The reconstructed images generated by EOTRN effectively highlight the detailed information and accurately reflect the health of the fundus. Moreover， EOTRN has fewer parameters， making it suitable for the deployment of embedded systems to realize real-time ultra-clear reconstruction of eye OCT images.

Abstract:To capture the three-dimensional irregularities of nodules in lung CT images and improve their diagnostic accuracy， a double-dimensional convolutional neural network （d3D-CNN） nodule diagnosis model from sieve to diagnosis is designed. Firstly， a lightweight 3D-CNN network is constructed， it is combined with full convolution operation， and the high optimization of convolution operation is used to complete nodule screening and generate suspected regions. Then， the space-slice attention mechanism is used to automatically learn the offset of the suspected region in space and slice sequence， design a deformable 3D convolution module， and combine it with ResNet101 to construct a high-precision 3D-CNN nodule diagnosis network for the final judgment of the screened suspected region. The comparative experimental results show that the recall rate of the proposed model reaches 88.9% under the false alarm rate of 1， which effectively improves the accuracy of benign and malignant diagnoses of pulmonary nodules.

Abstract:To solve the problem of low recognition rates on helmet dataset， this paper proposes a detection method based on an improved CenterNet network structure. To tackle the problem of poor prediction results in the multi-class classification of CenterNet， this paper attempts to improve the loss function. Therefore， Focal-Mse-One loss and Focal-Mse-Guss loss are proposed and compared with the original Focal loss； Aiming at the problem of low reusability of feature map in the reasoning process of CenterNet， ASFF and DASFF structures are proposed and compared. The experimental results show that the reasoning speed can reach 20.78 frames on GeForce GTX 1050 graphics card. When the IOU is 0.5， the mAP can reach 81.43% on the helmet dataset， which is 3.63% higher than the original CenterNet’s mAP. The improved method proposed in this paper can significantly improve the detec tion accuracy of safety helmet without a significant increase in reasoning time.

Abstract:As the feature size of integrated circuits approaches the physical limit， through-silicon-via-based three-dimensional integrated circuits （3D ICs） have become a trend to continue Moore’s Law. However， existing EDA tools， technology libraries and design methodologies are far from mature enough to achieve timing convergence of ultra-large-size interposers of 3D ICs. To address this issue， a new implementation flow for physical design of TSV-based 3D ICs using conventional EDA tools is proposed. Firstly， a thermal stress model is employed to project the silicon vias into 2D blockages， thereby dividing the entire 3D IC into several 2D ICs with blockages. Each of these 2D ICs can be implemented by traditional EDA tools， respectively. Secondly， to address the timing convergence difficulties of ultra-large-size interposers， this paper puts forward a new method， which first creates a couple of bounds throughout the layout and then iteratively moves pipeline cells affecting timing greatly between the bounds. Cells in bounds are not permitted to move during placement. This approach ensures a more organized initialization and reduces disorder， thus enabling convergence to be achieved. The whole flow is applied to the physical implementation of a practical 3D integrated circuit. The experimental results show that the proposed flow can optimize both the worst negative slack and the total negative slack by more than 98% compared with the original flow. Consequently， timing convergence is accomplished， and the feasibility of the proposed design flow is proved.

Abstract:A low-latency high-throughput Dynamic Virtual Output Queues Router for On-chip interconnect networks is proposed in this paper， which can reduce the router latency to two cycles by leveraging look-ahead routing computation and virtual output queues scheme. The simulation results show that， compared with the wormhole router and virtual-channel router， the network throughput on a 4×4 mesh increases by up to 46.9% and 28.6%， respectively， and outperforms doubled buffer virtual channel by 1.9% under the same input speedup. Under random synthetic traffic， the zero-load-latency of the network on chip is also reduced by 25.6% and 41%， respectively. Synthesis results indicate the frequency of router can reach 2.5 GHz.

Abstract:With the advancement of integrated circuit manufacturing technology and the improvement of chip integration，the demand for low-power chips has been steadily increasing. The clock network is responsible for more than 40% of the total power consumption of the chip. Consequently， optimizing the power consumption of the clock network has become one of the most important goals in the design of high-performance integrated circuits. In this paper， a new register clustering method is proposed to generate the leaf level topology of the clock tree. By carefully limiting the fan-out， load， and range of the clusters to reasonably group the registers， the method effectively reduces the number of buffer insertions and the total wiring length， and the clock network power consumption is also significantly reduced. The method is integrated into the traditional clock tree synthesis （CTS） flow， and its effectiveness is tested and analyzed on the ISCAS89 benchmark circuit. Experimental results show that the register clustering method effectively reduces the power dissipation of the clock network by more than 20% and the clock offset by more than 20%， without affecting the maximum delay of the clock tree.

Abstract:In the environment of energy saving and fierce peer competition， it is very urgent to optimize the power and performance optimization of application server clusters. Aiming at the deficiencies of the existing research in performance indicators and real-time performance， a real-time optimization scheme of cluster power and performance was proposed. This scheme combined the linear weighting method and the master objective method to optimize the cluster power and request drop rate， so converting the bi-objective optimization into a single-objective constrainted optimization. Firstly， based on the server load-power model in the CPU frequency equivalent continuous adjustment mode， the cluster optimization was described as a mixed integer quadratic programming problem by defining few variables. Then， variable splitting and variable conversion were used to transform the problem into a MILP （mixed integer linear programming） problem， and we introduced an SOS （special-Ordered set） constraint. Finally， the Gurobi optimizer was used to solve the MILP problem. Through further optimization of CPU frequency adjustment， the switching of CPU frequency was greatly reduced. Tests in various scenarios showed that the average solution time of the scheme was approximately 10 ms and the introduction of SOS constraint made the solution time more stable， which can ensure the real-time optimization.

Abstract:Focusing on the sparse channel estimation of orthogonal frequency division multiplexing （OFDM） modulation system using greedy pursuit algorithms， the problem of recovery performance degradation caused by wrong selection of atom is studied. Based on the analysis of the least-squares reconstruction process in the greedy pursuit algorithm， it is found that there is a severe overestimation of the atom coefficients on the wrong atoms. On this basis， a coefficient optimization scheme （COS） using the channel path delay correlation between adjacent symbols is proposed to improve the accuracy of sparse channel estimation when the atom selection is wrong. The simulation results show that the combining COS with the traditional orthogonal matching pursuit （OMP） algorithm and sparsity adaptive matching pursuit （SAMP） algorithm can effectively suppress the effect of wrong selection of atom on sparse channel estimation， and increase the estimation performance of greedy pursuit algorithms under low signal-to-noise ratio. The simulation tests under different multipath channel models show great robustness.

Abstract:To solve the problems of massive and heterogeneous data in the tobacco supply chain， high information barriers between enterprises， low trustworthiness of traceability data， and lack of supervision， a tobacco blockchain multi-chain traceability system was built using Hyperledger Fabric. This system implemented a multi-chain data storage and supervision model， enabling traceable data sharing and privacy data isolation through its multi-chain structure. Raft consensus algorithm was applied to develop an endorsement strategy， while smart contracts were designed to achieve the traceability for consumers and precise supervision for regulators. At the same time， a decentralized application （DAPP） for the traceability system was developed， ensuring user-friendly operations for consumers. To verify the feasibility of the traceability system， Caliper was used to conduct performance tests on the blockchain multi-chain traceability system. In terms of network performance， the blockchain network throughput was stable at 150 tps. In terms of trustworthiness， the blockchain transaction success rate was 100%. As for smart contract efficiency， the write throughput was stable at 150 tps， with the minimum query latency of 0.01 s and an average latency of 0.026 s. The results proved that the system met the performance requirements of practical application and was capable to realize the needs of anti-counterfeit traceability for consumers and accurate supervision for regulatory authorities.

Abstract:Aiming at the problems of current farmland sensor nodes information query and ontology knowledge construction， this paper conducts an analysis of the farmland moisture collection system. It proposes a lightweight semantic sensor network ontology and extracts the relationships from farmland information data. Additionally， it designs the FSSN ontology annotation method and utilizes Tf-Idf algorithm to analyze the semantic weight of farmland Ontology. FSSN-SDRM algorithm is proposed to construct the farmland moisture lightweight ontology model， analyze the growing environment of crops， set suitability reasoning rules， and the Jena API is employed for reasoning the annotated ontology model. According to the soil moisture information collected by the farmland sensor nodes， experiments are conducted on the lightweight middleware NVIDIA TX2 platform to query the correctness of reasoning information in the database， and compare the response time of the equipment. The experimental results show that lightweight FSSN annotation ontology compresses the average response time to 81 ms， which is 41.4% shorter than the uncommented ontology， and the time in TX2 platform is 12.81% shorter than that of the host side. The ontology model can quickly and accurately judge the suitability of crop growth environment and provides new ideas for agricultural information production.

Abstract:To solve simultaneous pickup and delivery among multiple demand points， a vehicle routing model that incorporates demand splitting and transfer is established. The constraint of dynamic variation of vehicle load， the constraint of multiple node access and the constraint of demand split transport are added in the model to improve the universality of the problem. In the optimization algorithm of the model， a hybrid approach combining arithmetic and ant colony optimization algorithm is employed to solve the problem. The algorithm follows a nested optimization structure， where the outer arithmetic optimization algorithm gets the task quantity of the delivery vehicle. The inner ant colony algorithm then optimizes the path， and provides feedback to the outer algorithm to continue to update and solve until the termination condition is met. At the same time， several enhancements are introduced to the hybrid arithmetic ant colony algorithm， such as incorporating probability coefficient， adding operator position update formula and updating dynamic tabu matrix. These enhancements aim to increase the diversity of solutions and improve the efficiency of the algorithm. Finally， the improved algorithm is verified by an example and compared with the hybrid whale algorithm and other algorithms to solve the problem in this paper.

Abstract:Event detection is one of the most important tasks in the field of natural language processing （NLP）. Its result is the key information supporting downstream tasks， such as information extraction， text classification and event reasoning. BERT model has achieved remarkable achievements in event detection. However， it cannot effectively obtain long-distance and structured text information. To alleviate this problem， feedback-based GCNs network is proposed to capture text structure information in this paper， and it can solve the problem of semantic information attenuation caused by GCNs. This paper first uses BERT model to obtain semantic features of the text， then adopts GCNs integrated into the feedback network to extract the syntactic structure features of the text， and finally employs multiple classifiers to identify and classify event trigger words. The experimental results on the open dataset ACE 2005 show that the F1 value of the event detection method proposed in the task of event trigger word recognition and classification has reached 74.46% and 79.49%， respectively， which gains an average increase of 4.13% and 4.79% compared with the existing work.

Abstract:Script Event Prediction refers to predicting the subsequent event based on a given existing chain of context events. In the real world， the relationship of different events can be naturally represented as a graph structure， where events serve as nodes and their temporal or causal relations are depicted as edges. However， previous approaches that automatically constructed event graphs suffer from sparsity problem due to the limited scale of corpus and the incapability of information extraction tools. Moreover， they fail to integrate information from higher order nodes to support multi-step reasoning. To remedy this， we propose a Dense Event Graph （DEG） approach which use a learnable multi-dimensional weighted adjacency matrix to address the sparsity issue and characterize the relation strengths between events. To embed the DEG， we propose a general framework capable of combining high-order event evolution information into the event representations. Experimental results on the multiple choice narrative cloze （MCNC） and coherent multiple choice narrative cloze （CMCNC） demonstrate the effectiveness of our approach.