For the shortcomings of poor interpretation ability and instability in neural network training， a Chambolle- Pock （CP） algorithm optimized denoising network based on Total Variational （TV） regularization， CPTV-Net， was proposed to solve the denoising problem of Low-Dose Computed Tomography （LDCT） images. Firstly， the TV constraint term was introduced into the L1 regularization term model to preserve the structural information of the image. Secondly， the CP algorithm was used to solve the denoising model and obtain specific iterative steps to ensure the convergence of the algorithm. Finally， the shallow CNN （Convolutional Neural Network） was used to learn the iterative formula of the primal dual variables of the linear operation. The neural network was used to calculate the solution of the model， and the network parameters were collected to optimize the combined data. The experimental results on simulated and real LDCT datasets show that compared with five advanced denoising methods such as REDCNN （Residual Encoder-Decoder Convolutional Neural Network） and TED-Net （Transformer Encoder-decoder Dilation Network）， CPTV-Net has the best Peak Signal-to-Noise Ratio （PSNR）， Structural SIMilarity （SSIM）， and Visual Information Fidelity （VIF） evaluation values， and can generate LDCT images with significant denoising effect and the most details preserved.
Aiming at the high computational complexity and large memory consumption of the existing super-resolution reconstruction networks， a lightweight image super-resolution reconstruction network based on Transformer-CNN was proposed， which made the super-resolution reconstruction network more suitable to be applied on embedded terminals such as mobile platforms. Firstly， a hybrid block based on Transformer-CNN was proposed， which enhanced the ability of the network to capture local-global depth features. Then， a modified inverted residual block， with special attention to the characteristics of the high-frequency region， was designed， so that the improvement of feature extraction ability and reduction of inference time were realized. Finally， after exploring the best options for activation function， the GELU （Gaussian Error Linear Unit） activation function was adopted to further improve the network performance. Experimental results show that the proposed network can achieve a good balance between image super-resolution performance and network complexity， and reaches inference speed of 91 frame/s on the benchmark dataset Urban100 with scale factor of 4， which is 11 times faster than the excellent network called SwinIR （Image Restoration using Swin transformer）， indicates that the proposed network can efficiently reconstruct the textures and details of the image and reduce a significant amount of inference time.
Federated Learning （FL） allows users to share and interact with multiple parties without directly uploading the original data， effectively reducing the risk of privacy leaks. However， existing research suggests that the adversary can still reconstruct raw data through shared gradient information. To further protect the privacy of federated learning， a deep shadow defense scheme of federated learning based on Generative Adversarial Network （GAN） was proposed. The original real data distribution features were learned by GAN and replaceable shadow data was generated. Then， the original model trained on real data was replaced by a shadow model trained on shadow data and was not directly accessible to the adversary. Finally， the real gradient was replaced by the shadow gradient generated by the shadow data in the shadow model and was not accessible to the adversary. Experiments were conducted on CIFAR10 and CIFAR100 datasets for comparison of the proposed scheme with the five defense schemes of adding noise， gradient clipping， gradient compression， representation perturbation and local regularization and sparsification. On CIFAR10 dataset， the Mean Square Error （MSE） and the Feature Mean Square Error （FMSE） of the proposed scheme were 1.18-5.34 and 4.46-1.03×107 times， and the Peak Signal-to-Noise Ratio （PSNR） of the proposed scheme was 49.9%-90.8%. On CIFAR100 dataset， the MSE and the FMSE of the proposed scheme were 1.04-1.06 and 5.93-4.24×103 times， and the PSNR of the proposed scheme was 96.0%-97.6%. Compared with the deep shadow defense method， the proposed scheme takes into account the actual attack capability of the adversary and the problems in shadow model training， and designs threat models and shadow model generation algorithms. It performs better in theory analysis and experiment result that of the comparsion schemes， and it can effectively reduce the risk of federated learning privacy leaks while ensuring accuracy.
In order to solve the problem of most current research work only considering the factors affecting pedestrian interaction， based on State-Refinement Long Short-Term Memory （SR-LSTM） and attention mechanism， a Social-Interaction Generative Adversarial Network （SIGAN） for pedestrian trajectory prediction was proposed， namely SRA-SIGAN， where GAN was utilized to learn movement patterns of target pedestrians. Firstly， SR-LSTM was used as a location encoder to extract the information of motion intention. Secondly， the influence of pedestrians in the same scene was reasonably assigned by setting the velocity attention mechanism， thereby handling the pedestrian interaction better. Finally， the predicted future trajectory was generated by the decoder. Experimental results on several public datasets show that the performance of SRA-SIGAN model is good on the whole. Specifically on the Zara1 dataset， compared with SR-LSTM model，the Average Displacement Error （ADE）and Final Displacement Error （FDE）of SRA-SIGAN were reduced by 20.0% and 10.5%，respectively；compared with the SIGAN model，the ADE and FDE of SRA-SIGAN were decreased by 31.7% and 24.4%，respectively.
In certain fixed industrial application scenarios， the tolerance of the target detection algorithms to miss detection is very low. However， while increasing the recall， some non-overlapping virtual frames are likely to be regularly generated around the target. The traditional Non-Maximum Suppression （NMS） strategy has the main function to suppress multiple repeated detection frames of the same target， and cannot solve the above problem. To this end， an anisotropic NMS method was designed by adopting different suppression strategies for different directions around the target， and was able to effectively eliminate the regular virtual frames. The target shape and the regular virtual frame in a fixed industrial scene often have a certain relevance. In order to promote the accurate execution of anisotropic NMS in different directions， a ratio Intersection over Union （IoU） loss function was designed to guide the model to fit the shape of the target. In addition， an automatic labeling dataset augmentation method was used for the regular target， which reduced the workload of manual labeling and enlarged the scale of the dataset. Experimental results show that the proposed method has significant effects on the roll groove detection dataset， and when it is applied to the YOLO （You Only Look Once） series of algorithms， the detection precision is improved without reducing the speed. At present， the algorithm has been successfully applied to the production line of a cold rolling mill that automatically grabs rolls.
Heuristic and machine learning based code smell detection methods have been proved to have limitations， and most of these methods focus on the common code smells. In order to solve these problems， a deep learning based method was proposed to detect three relatively rare code smells which are related to coupling， those are Intensive Coupling， Dispersed Coupling and Shotgun Surgery. First， the metrics of three code smells were extracted， and the obtained data were processed. Second， a deep learning model combining Convolutional Neural Network （CNN） and attention mechanism was constructed， and the introduced attention mechanism was able to assign weights to the metric features. The datasets were extracted from 21 open source projects， and the detection methods were validated in 10 open source projects and compared with CNN model. Experimental results show that the proposed model achieves the better performance with the code smell precisions of 93.61% and 99.76% for Intensive Coupling and Dispersed Coupling respectively， and the CNN model achieves the better results with the code smell precision of 98.59% for Shotgun Surgery.
A multimodal sequential recommendation algorithm based on contrastive learning technology was proposed to improve the accuracy of sequential recommendation algorithm by using multimodal information of commodities. Firstly， to obtain the visual representations such as the color and shape of the product， the visual modal information of the product was extracted by utilizing the contrastive learning framework， where the data enhancement was performed by changing the color and intercepting the center area of the product. Secondly， the textual information of each commodity was embedded into a low-dimensional space， so that the complete multimodal representation of each commodity could be obtained. Finally， a Recurrent Neural Network （RNN） was used for modeling the sequential interactions of multimodal information according to the time sequence of the product， then the preference representation of user was obtained and used for commodity recommendation. The proposed algorithm was tested on two public datasets and compared with the existing sequential recommendation algorithm LESSR. Experimental results prove that the ranking performance of the proposed algorithm is improved， and the recommendation performance remains basically unchanged after the feature dimension value reaches 50.
Because the traditional developer recommendation methods focus on analyzing the developers’ professional abilities and the interaction information with the tasks， without considering the problem of collaboration between the developers， a developer recommendation method based on Environment-Class， Agent， Role， Group， and Object （E-CARGO） model was proposed. Firstly， the developer collaborative development process was described as a role-based collaboration problem and modeled by E-CARGO model combining the characteristics of collaborative development. Then， a fuzzy judgment matrix was established by Fuzzy Analytic Hierarchy Process （FAHP） method to obtain the developer ability index weights and weighted sum of them， thereby obtaining the set of historical comprehensive ability evaluation of the developers. Finally， in view of the uncertainty and dynamic characteristics of the developers’ comprehensive ability evaluation， the cloud model theory was used to analyze the set of historical comprehensive ability evaluation of the developers to obtain the developers’ competence for each task， and the cplex optimization package was used to solve the developer recommendation problem. Experimental results show that the proposed method can obtain the best developer recommendation results within an acceptable time range， which verifies the effectiveness of the proposed method.
Query is the main workload of a database system， and its efficiency determines the performance of the database system. There are multiple execution plans for a query， and the existing query optimizers can only statically select a better execution plan for a query according to the configuration parameters of the database system. There are complex and variable resource contentions between concurrent queries， and such contentions are difficult to be reflected accurately through configuration parameters； besides， the efficiency of the same execution plan is not consistent in different scenarios. The selection of the execution plans for concurrent queries needs to consider the influence between queries — query interaction. Based on the above， a metric for measuring the influence of query interaction on the query under concurrent query called QIs （Query Interactions） was proposed. For the selection of query execution plan under concurrent query， a method called TRating （Time Rating） was proposed to dynamically select the execution plan for the query. In the method， the influence of query interaction on the queries executed with different plans in the query combination was measured， and the plan with small influence of query interaction was selected as the better execution plan for the query. Experimental results show that TRating can select a better execution plan for the query with an accuracy of 61%， which is 25% higher than that of the query optimizer； and the accuracy of the proposed method is as high as 69% when selecting suboptimal execution plan for the query.
