Due to the high complexity， weak individual differences， and short inter-class distances of fundus image features， pure Convolutional Neural Networks （CNNs） and attention based networks cannot achieve satisfactory accuracy in fundus disease image classification tasks. To this end， Attention Bilinear Involution Neural Network （ABINN） model was implemented for fundus disease image classification by using the involution operator. The parameter amount of ABINN model was only 11% of that of the traditional Bilinear Convolutional Neural Network （BCNN） model. In ABINN model， the underlying semantic information and spatial structure information of the fundus image were extracted and the second-order features of them were fused. It is an effective parallel connection between CNN and attention method. In addition， two instantiation methods for attention calculation based on involution operator， Attention Subnetwork based on PaTch （AST） and Attention Subnetwork based on PiXel （ASX）， were proposed. These two methods were able to calculate attention within the CNN basic structure， thereby enabling bilinear sub-networks to be trained and fused in the same architecture. Experimental results on public fundus image dataset OIA-ODIR show that ABINN model has the accuracy of 85%， which is 15.8 percentage points higher than that of the common BCNN model and 0.9 percentage points higher than that of TransEye （Transformer Eye） model.

Aiming at the low efficiency and the complex simulation process of the traditional reconstruction methods for porous media such as Multi-Point Statistics(MPS)which require scanning the training image many times and to obtain simulation results by complex probability calculations,a method to reconstruct porous media using adaptive deep transfer learning was presented. Firstly,deep neural network was used to extract the complex features from the training image of porous media. Secondly,the adaptive layer was added in deep transfer learning to reduce the difference in data distribution between training data and prediction data. Finally,through copying features by transfer learning,the simulation result consistent with the real training data was obtained. The performance of the proposed method was evaluated by comparing with the classical porous media reconstruction method MPS in multiple-point connectivity curve,variogram curve and porosity. The results indicate that the proposed method has high reconstruction quality. Meanwhile,the method has the average running time reduced from 840 s to 166 s,the average CPU usage dropped from 98% to 20%,and the average memory utilization decreased by 69%. The proposed method significantly improves the efficiency of porous media reconstruction under the premise of ensuring better quality of reconstruction results.

Item-based collaborative filtering learns user preferences from the user's historical interaction items and recommends similar new items based on the user's preferences. The existing collaborative filtering methods assume that a set of historical items that user has interacted with have the same impact on user, and all historical interaction items are considered to have the same contribution to the prediction of target item, which limits the accuracy of these recommendation methods. In order to solve the problems, a new collaborative filtering recommendation algorithm based on dual most relevant attention network was proposed, which contained two attention network layers. Firstly, the item-level attention network was used to assign different weights to different historical items in order to capture the most relevant items in the user historical interaction items. Then, the item-interaction-level attention network was used to perceive the correlation degrees of the interactions between the different historical items and the target item. Finally, the fine-grained preferences of users on the historical interaction items and the target item were simultaneously captured through the two attention network layers, so as to make the better recommendations for the next step. The experiments were conducted on two real datasets of MovieLens and Pinterest. Experimental results show that, the proposed algorithm improves the recommendation hit rate by 2.3 percentage points and 1.5 percentage points respectively compared with the benchmark model Deep Item-based Collaborative Filtering (DeepICF) algorithm, which verifies the effectiveness of the proposed algorithm on making personalized recommendations for users.

Concerning the problems that the traditional Convolutional Neural Network (CNN) ignores the contextual semantics of words when performing sentiment analysis tasks and CNN loses a lot of feature information during max pooling operation at the pooling layer, which limit the text classification performance of model, a parallel hybrid neural network model, namely CA-BGA (Convolutional Neural Network Attention and Bidirectional Gated Recurrent Unit Attention), was proposed. Firstly, a feature fusion method was adopted to integrate Bidirectional Gated Recurrent Unit (BiGRU) into the output of CNN, thus semantic learning was enhanced by integrating the global semantic features of sentences. Then, the attention mechanism was introduced between the convolutional layer and the pooling layer of CNN and at the output of BiGRU to reduce noise interference while retaining more feature information. Finally, a parallel hybrid neural network model was constructed based on the above two improvement strategies. Experimental results show that the proposed hybrid neural network model has the characteristic of fast convergence, and effectively improves the F1 value of text classification. The proposed model has excellent performance in Chinese short text sentiment analysis tasks.

The physical layer secure transmission in two-way multi-relay system can not obtain the accurate Channel State Information (CSI) of eavesdroppers. In order to solve the problem, a robust joint physical layer secure transmission scheme of multi-relay cooperative beamforming and artificial noise was proposed to maximize the secrecy sum rate in the worst case of channel state under the total power constraint of system. In the proposed scheme, the problem to be solved was a complex non-convex optimization problem. The alternating iteration and Successive Convex Approximation (SCA) methods were used for the alternating optimization iteration of beamforming vector, artificial noise covariance matrix and source node transmit power, and the optimal solution of the above problem was obtained. The simulation results verify the effectiveness of the proposed scheme and show that the proposed scheme has better security performance.

In order to reduce the entries of routing tables and avoid using large numbers of Virtual Channels (VC), a Region-based Fault Tolerant Routing (RFTR) algorithm was proposed for wormhole switching 2D Mesh Network on Chip (NoC) to reduce the amount of hardware resources. According to the positions of faulty nodes and links, the 2D Mesh network was divided into several rectangular regions. Within each region the packet could be routed by deterministic or adaptive routing algorithms, while among these regions the routing path was determined by up ^*/down ^* routing algorithm. Besides, with the Channel Dependency Graph (CDG) model, the proposed algorithm was proved to be deadlock-free using only two VCs. In a 6×6 Mesh network, the RFTR algorithm can reduce the amount of routing table resources by 25%. Simulation results show that, with the same amount of buffer resources, the RFTR algorithm can achieve an equivalent or even higher performance compared to up ^*/down ^* and segment-based routing algorithms.

To keep the trade-off of time complexity and accuracy of community detection in complex networks, Community Detection Algorithm based on Clustering Granulation (CGCDA) was proposed in this paper. The granules were regarded as communities so that the granulation for a network was actually the community partition of a network. Firstly, each node in the network was regarded as an original granule, then the granule set was obtained by the initial granulation operation. Secondly, granules in this set which satisfied granulation coefficient were merged by clustering granulation operation. The process was finished until granulation coefficient was not satisfied in the granule set. Finally, overlapping nodes among some granules were regard as isolated points, and they were merged into corresponding granules based on neighbor nodes voting algorithm to realize the community partition of complex network. Newman Fast Algorithm (NFA), Label Propagation Algorithm (LPA), CGCDA were realized on four benchmark datasets. The experimental results show that CGCDA can achieve modularity 7.6% higher than LPA and time 96% less than NFA averagely. CGCDA has lower time complexity and higher modularity. The balance between time complexity and accuracy of community detection is achieved. Compared with NFA and LPA, the whole performance of CGCDA is better.

As to the problems that Millimeter Wave (MMW) image is contaminated by much unknown noise and has lower resolution, and considering the non-linear filter property of Fuzzy Radial Basis Function Neural Network (F-RBFNN) and the self-adaptive denoising property of Sparse Representation (SR) based on K-Singular Value Decomposition (K-SVD), a MMW restoration method was proposed by combining F-RBFNN and sparse representation. In F-RBFNN, the knowledge expression of fuzzy logic and the reasoning ability were combined with the RBFNN's capabilities of fast learning and generalization. In order to realize the non-linear filtering to the MMW image, F-RBFNN's structure and parameters were adjusted according to the real problem. Furthermore, utilizing the advantages of sparse representation method, which the sparse representation behaves the visual characteristic and can denoise effectively when maintaining features of the object, the training results of F-RBFNN were locally denoised once again, and the MMW image with high resolution was obtained. Using the Relative Single Noise Ratio (RSNR) criterion to measure the quality of denoised images, the simulation results show that, compared with other denoising methods such as F-RBFNN, K-SVD denoising, and wavelet denoising, the proposed method combining F-RBFNN and SR can better restore the quality of MMW image.

Error Correcting Output Codes (ECOC) is an effective algorithm to handle multi-class problem; however, the ECOC coding is only on the class level and the ECOC matrix is pre-designed. A novel classification algorithm based on hierarchical ECOC was proposed. The algorithm first used KNNModel to build multiple clusters on a given dataset and chose few clusters for each class as representatives to construct a hieratical coding matrix in training phase, and then the matrix was used to train each single classifier. In testing phase, the proposed method makes the most of the merits of KNNModel and ECOC through models combination. Experimental results in the UCI data sets show the effectiveness of the proposed method.