In order to solve the problems， such as insufficient search ability and low search efficiency of Heap-Based optimizer （HBO） in solving complex problems， a Differential disturbed HBO （DDHBO） was proposed. Firstly， a random differential disturbance strategy was proposed to update the best individual’s position to solve the problem of low search efficiency caused by not updating of this individual by HBO. Secondly， a best worst differential disturbance strategy was used to update the worst individual’s position and strengthen its search ability. Thirdly， the ordinary individual’s position was updated by a multi-level differential disturbance strategy to strengthen information communication among individuals between multiple levels and improve the search ability. Finally， a dimension-based differential disturbance strategy was proposed for other individuals to improve the probability of obtaining effective solutions in initial stage of original updating model. Experimental results on a large number of complex functions from CEC2017 show that compared with HBO， DDHBO has better optimization performance on 96.67% functions and less average running time （3.445 0 s）， and compared with other state-of-the-art algorithms， such as Worst opposition learning and Random-scaled differential mutation Biogeography-Based Optimization （WRBBO）， Differential Evolution and Biogeography-Based Optimization （DEBBO）， Hybrid Particle Swarm Optimization and Grey Wolf Optimizer （HGWOP）， etc.， DDHBO also has significant advantages.

In current international society， as the international language， English characters appear in many public occasions， as well as the Chinese pinyin characters in Chinese environment. When these characters appear in the image， especially in the image with complex style， it is difficult to edit and modify them directly. In order to solve the problems， an image character editing method based on improved character generation network named Font Adaptive Neural network （FANnet） was proposed. Firstly， the salience detection algorithm based on Histogram Contrast （HC） was used to improve the Character Adaptive Detection （CAD） model to accurately extract the image characters selected by the user. Secondly， the binary image of the target character that was almost consistent with the font of the source character was generated by using FANnet. Then， the color of source characters were transferred to target characters effectively by the proposed Colors Distribute-based Local （CDL） transfer model based on color complexity discrimination. Finally， the target editable characters that were highly consistent with the font structure and color change of the source character were generated， so as to achieve the purpose of character editing. Experimental results show that， on MSRA-TD500， COCO-Text and ICDAR datasets， the average values of Structural SIMilarity（SSIM）， Peak Signal-to-Noise Ratio （PSNR） and Normalized Root Mean Square Error （NRMSE） of the proposed method are 0.776 5， 18.321 1 dB and 0.435 8 respectively， which are increased by 18.59%，14.02% and decreased by 2.97% comparing with those of Scene Text Editor using Font Adaptive Neural Network（STEFANN） algorithm respectively， and increased by 30.24%，23.92% and decreased by 4.68% comparing with those of multi-modal few-shot font style transfer model named Multi-Content GAN（MC-GAN） algorithm（with 1 input character）respectively. For the image characters with complex font structure and color gradient distribution in real scene， the editing effect of the proposed method is also good. The proposed method can be applied to image reuse， image character computer automatic error correction and image text information restorage.

In order to solve the problems of poor denoising effect and low computational efficiency in image denoising, a double blind denoising algorithm based on Denoising Convolutional Neural Network (DnCNN) and Conditional Generative Adversarial Net (CGAN) was proposed. Firstly, the improved DnCNN model was used as the CGAN generator to capture the noise distribution of the noisy image. Secondly, the noisy image after eliminating the noise distribution and the tag were sent to the discriminator to distinguish the noise reduction image. Thirdly, the results of discrimination were used to optimize the hidden layer parameters of the whole model. Finally, a balance between the generator and the discriminator was achieved in the game, and the generator's residual capture ability was optimal. Experimental results show that on Set12 dataset, when the noise levels are 15, 25, 50 respectively:compared with the DnCNN algorithm, the proposed algorithm has the Peak Signal-to-Noise Ratio (PSNR) increased by 1.388 dB, 1.725 dB and 1.639 dB respectively based on the error evaluation index between pixel points. Compared with the existing algorithms such as Block Matching 3D (BM3D), Weighted Nuclear Norm Minimization (WNNM), DnCNN, Cascade of Shrinkage Fields (CSF) and ConSensus neural NETwork (CSNET), the proposed algorithm has the index value of Structural SIMilarity (SSIM) improved by 0.000 2 to 0.104 1 on average based on the evaluation index of structural similarity. The above experimental results verify the superiority of the proposed algorithm.

Concerning the the size and morphological diversity of thyroid tissue and the complexity of surrounding tissue in thyroid ultrasound images, an ultrasound thyroid segmentation network based on feature fusion and dynamic multi-scale dilated convolution was proposed. Firstly, the dilated convolutions with different dilation rates and dynamic filters were used to fuse the global semantic information of different receptive domains and the semantic information in the context details with different ranges, so as to improve the adaptability and accuracy of the network to multi-scale targets. Then, the hybrid upsampling method was used to enhance the spatial information of high-dimensional semantic features and the context information of low-dimensional spatial features during feature dimensionality reduction. Finally, the spatial attention mechanism was introduced to optimize the low-dimensional features of the image, and the method of fusing high- and low-dimensional features was applied to retain the useful features of high- and low-dimensional feature information with the elimination of the redundant information and improve the network's ability to distinguish the background and foreground of the image. Experimental results show that the proposed method has an accuracy rate of 0.963±0.026, a recall rate of 0.84±0.03 and a dice coefficient of 0.79±0.03 in the public dataset of thyroid ultrasound images. It can be seen that the proposed method can solve the problems of large difference of tissue morphology and complex surrounding tissues.

Concerning the problems of the lack of a large number of relationships in the current Knowledge Graph (KG), and the lack of full consideration of the hidden information in the multi-hop path between two entities when performing relationship reasoning, a relationship reasoning method combining multi-hop relationship path information was proposed. Firstly, for the given candidate relationships and two entities, the convolution operation was used to encode the multi-hop relationship path connecting the two entities into a low-dimensional space and extract the information. Secondly, the Bidirectional Long Short-Term Memory (BiLSTM) network was used for modeling to generate the relationship path representation vector, and the attention mechanism was used to combine it with the candidate relationship representation vector. Finally, a multi-step reasoning method was used to find the relationship with the highest matching degree as the reasoning result and judge its precision. Compared with the current popular Path Ranking Algorithm (PRA), the neural network model named Path-RNN and reinforcement learning model named MINERVA, the proposed algorithm had the Mean Average Precision (MAP) increased by 1.96,8.6 and 1.6 percentage points respectively when using the large-scale knowledge graph dataset NELL995 for experiments. And when using the small-scale knowledge graph dataset Kinship for experiments, the proposed algorithm had the MAP improved by 21.3,13 and 12.1 percentage points respectively compared to PRA and MINERVA. The experimental results show that the proposed method can infer the relationship links between entities more accurately.

The Closed-Loop Layout Problem (CLLP) is an NP-hard mixed optimization problem, in which an optimal placement order of facilities is found along adjustable rectangle loop with the objection of minimizing the total transport cost of material flow between facilities. In most of the existing methods, meta-heuristic algorithm was used to find the optimal order for the placement of facilities, and enumeration method was applied to find the optimal size of the rectangle loop, which causes extremely low efficiency. To solve this problem, a Hybrid Population-Based Incremental Learning (HPBIL) algorithm was proposed for solving CLLP. In the algorithm, the Discrete Population-Based Incremental Learning (DPBIL) operator and Continuous PBIL (CPBIL) operator were used separately to search the optimal placement order of facilities and the size of rectangle loop at the same time, which improved the efficiency of search. Furthermore, a local search algorithm was designed to optimize some good solutions in each iteration, enhancing the refinement ability. Simulation experiments were carried out on 13 CLLP instances. The results show that HPBIL algorithm finds the best new optimal layouts on 9 instances, and is significantly superior to the algorithms to be compared on the optimization ability for CLLP.

Block Diagonal Representation (BDR) model can efficiently cluster data by using linear representation, but it cannot make good use of non-linear manifold information commonly appeared in high-dimensional data. To solve this problem, the improved Block Diagonal Representation based on Neighbor Graph (BDRNG) clustering algorithm was proposed to perform the linear fitting of the local geometric structure by the neighbor graph and generate the block-diagonal structure by using the block-diagonal regularization. In BDRNG algorithm, both global information and local data structure were learned at the same time to achieve a better clustering performance. Due to the fact that the model contains the neighbor graph and non-convex block-diagonal representation norm, the alternative minimization was adopted by BDRNG to optimize the solving algorithm. Experimental results show that:on the noise dataset, BDRNG can generate the stable coefficient matrix with block-diagonal form, which proves that BDRNG is robust to the noise data; on the standard datasets, BDRNG has better clustering performance than BDR, especially on the facial dataset, BDRNG has the clustering accuracy 8% higher than BDR.

In view of the problem that most existing network situation prediction methods are unable to mine the deep information in the data and need to manually extract and construct features, a deep feature network situation prediction method named DFS-Seq2Seq (Deep Feature Synthesis-Sequence to Sequence) was proposed. First, the data produced by network streams, logs and system events were cleaned, and the deep feature synthesis algorithm was used to automatically synthesize the deep relation features. Then the synthesized features were extracted by the AutoEncoder (AE). Finally, the data was estimated by using the Seq2Seq (Sequence to Sequence) model constructed by Long Short-Term Memory (LSTM). Through a well-designed experiment, the proposed method was verified on the public dataset Kent2016. Experimental results show that when the depth is 2, compared with four classification models including Support Vector Machine (SVM), Bayes, Random Forest (RF) and LSTM, the proposed method has the recall rate increased by 7.4%, 11.5%, 6.5% and 3.0%, respectively. It is verified that DFS-Seq2Seq can effectively identify dangerous events in network authentication and effectively predict network situation in practice.

Random multiple access communication is an indispensable part of computer communication research. A three-slot P-Persistent Carrier Sense Multiple Access (P-CSMA) protocol with variable collision duration in Wireless Sensor Network (WSN) was proposed to solve the problem of traditional P-CSMA protocol in transmitting and controlling WSN and energy consumption of system. In this protocol, the collision duration was added to the traditional two-time-slot P-CSMA protocol in order to change the system model to three-time-slot model, that is, the duration of information packet being sent successfully, the duration of packet collision and the idle duration of the system.Through the modeling, the throughput, collision rate and idle rate of the system under this model were analyzed. It was found that by changing the collision duration, the loss of the system was reduced. Compared with the traditional P-CSMA protocol, this protocol makes the system performance improved, and makes the lifetime of the system nodes obtained based on the battery model obviously extended. Through the analysis, the system simulation flowchart of this protocol is obtained. Finally, by comparing and analyzing the theoretical values and simulation values of different indexes, the correctness of the theoretical derivation is proved.

The data in internet social media has the characteristics of fast transmission, high user participation and complete coverage compared with traditional media under the background of the rise of various platforms on the internet.There are various topics that people pay attention to and publish comments in, and there may exist deeper and more fine-grained sub-topics in the related information of one topic. A survey of sub-topic detection based on internet social media, as a newly emerging and developing research field, was proposed. The method of obtaining topic and sub-topic information through social media and participating in the discussion is changing people’s lives in an all-round way. However, the technologies in this field are not mature at present, and the researches are still in the initial stage in China. Firstly, the development background and basic concept of the sub-topic detection in internet social media were described. Secondly, the sub-topic detection technologies were divided into seven categories, each of which was introduced, compared and summarized. Thirdly, the methods of sub-topic detection were divided into online and offline methods, and the two methods were compared, then the general technologies and the frequently used technologies of the two methods were listed. Finally, the current shortages and future development trends of this field were summarized.

Different clustering algorithms are used to design their own strategies. However,each technology has certain limitations when it executes a particular dataset. An adequate choice of Discrimination Information Method(DIM)can ensure the document clustering. To solve these problems,a DIM of Document Clustering based on Consensus and Classification (DCCC) was proposed. Firstly,Clustering by DIM (CDIM) was used to solve the generation of initial clustering for dataset,and two initial cluster sets were generated by two different CDIMs. Then,two initial cluster sets were initialized again by different parameter methods,and a consensus was established by using the relationship between the cluster label information,so as to maximize the sum of documents' discrimination number. Finally,Discrimination Text Weight Classification(DTWC)was chosen as text classifier to assign new cluster label to the consensus,the base partitions were altered by training the text classifier,and the final partition was obtained based on the predicted label information. Experiments on 8 network datasets for clustering verification by BCubed's precision and recall index were carried out. Experimental results show that the clustering results of the proposed consensus and classification method are superior to those of comparison methods.

In order to effectively solve the problems of not obvious boundaries, unclear abdominal organ display caused by high-frequency detail loss as well as the inconvenient application of single-model single-scale reconstruction in the super-resolution reconstruction of abdominal Magnetic Resonance Imaging (MRI) images, a multi-scale super-resolution algorithm based on parallel channel-spatial attention mechanism was proposed. Firstly, parallel channel-spatial attention residual blocks were built. The correlation between the key area and high-frequency information was obtained by the spatial attention module, and the channel attention module was used to study the weights of the channels of the image to the key information response degree. At the same time, the feature extraction layer of the network was widened to increase the feature information flowing into the attention module. In addition, the weight normalized layer was added to ensure the training efficiency of the network. Finally, a multi-scale up-sampling layer was applied at the end of the network to increase the flexibility and applicability of the network. Experimental results show that, compared with the image super-resolution using very deep Residual Channel Attention Network (RCAN), the proposed algorithm has the Peak Signal-to-Noise Ratio (PSNR) averagely increased by 0.68 dB at the×2,×3 and×4 scales. The proposed algorithm effectively improves the reconstructed image quality.

The traditional spectral clustering algorithm needs too much time to perform eigen-decomposition when the number of samples is very large. In order to solve the problem, a fast spectral clustering algorithm without eigen-decomposition was proposed to reduce the time overhead by multiplication update iteration. Firstly, the Nyström algorithm was used for random sampling in order to establish the relationship between the sampling matrix and the original matrix. Then, the indicator matrix was updated iteratively based on the principle of multiplication update iteration. Finally, the correctness and convergence analysis of the designed algorithm were given theoretically. The proposed algorithm was tested on five widely used real datasets and three synthetic datasets. Experimental results on real datasets show that:the average Normalized Mutual Information (NMI) of the proposed algorithm is 0.45, which is improved by 12.5% compared with that of the k-means clustering algorithm; the computing time of the proposed algorithm achieves 61.73 s, which is decreased by 61.13% compared with that of the traditional spectral clustering algorithm; and the performance of the proposed algorithm is superior to that of the hierarchical clustering algorithm, which verify the effectiveness of the proposed algorithm.

Aiming at how to merge the nodes' attributes and network structure information to realize the classification of social network nodes, a social network node classification algorithm based on graph encoder network was proposed. Firstly, the information of each node was propagated to its neighbors. Secondly, for each node, the possible implicit relationships between itself and its neighbor nodes were mined through neural network, and these relationships were merged together. Finally, the higher-level features of each node were extracted based on the information of the node itself and the relationships with the neighboring nodes and were used as the representation of the node, and the node was classified according to this representation. On the Weibo dataset, compared with DeepWalk model, logistic regression algorithm and the recently proposed graph convolutional network, the proposed algorithm has the classification accuracy greater than 8%; on the DBLP dataset, compared with multilayer perceptron, the classification accuracy of this algorithm is increased by 4.83%, and is increased by 0.91% compared with graph convolutional network.

In order to improve the quality of medical image super-resolution reconstruction, a wide residual super-resolution neural network algorithm based on depthwise separable convolution was proposed. Firstly, the depthwise separable convolution was used to improve the residual block of the network, widen the channel of the convolution layer in the residual block, and pass more feature information into the activation function, making the shallow low-level image features in the network easier transmitted to the upper level, so that the quality of medical image super-resolution reconstruction was enhanced. Then, the network was trained by group normalization, the channel dimension of the convolutional layer was divided into groups, and the normalized mean and variance were calculated in each group, which made the network training process converge faster, and solved the difficulty of network training because the depthwise separable convolution widens the number of channels. Meanwhile, the network showed better performance. The experimental results show that compared with the traditional nearest neighbor interpolation, bicubic interpolation super-resolution algorithm and the super-resolution algorithm based on sparse expression, the medical image reconstructed by the proposed algorithm has richer texture detail and more realistic visual effects. Compared with the super-resolution algorithm based on convolutional neural network, the super-resolution neural network algorithm based on wide residual and the generative adversarial-network super-resolution algorithm, the proposed algorithm has a significant improvement in PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural SIMilarity index).

When traditional signature algorithms such as blind signature and group signature applied to heterogeneous networks of blockchain, they might have problems like relying on trusted centers or low efficiency. Aiming at the problems, a threshold signature scheme suitable for blockchain electronic voting scenes was proposed. The proposed scheme was based on the Asmuth-Bloom secret sharing scheme and did not need a trusted center. Firstly, the signature was generated by the collaboration of blockchain nodes, implementing mutual verification between nodes and improving the node credibility. Secondly, a mechanism of nodes joining and exiting was established to adapt to the high mobility of the blockchain nodes. Finally, the node private keys were updated regularly to resist mobile attacks and make them forward-secure. Security analysis shows that the security of the scheme is based on the discrete logarithm problem, so that the scheme can effectively resist mobile attacks and is forward-secure. The performance analysis shows that compared with other schemes, this scheme has lower computational complexity in the signature generation and verification phases. The results show that the proposed scheme can be well applied to blockchain electronic voting scenes.

On-line detection of fabric defects is a major problem faced by textile industry. Aiming at the problems such as high false positive rate, high false negative rate and low real-time in the existing detection of fabric defects, an on-line detection algorithm for fabric defects based on deep learning was proposed. Firstly, based on GoogLeNet network architecture, and referring to classical algorithm of other classification models, a fabric defect classification model suitable for actual production environment was constructed. Secondly, a fabric defect database was set up by using different kinds of fabric pictures marked by quality inspectors, and the database was used to train the fabric defect classification model. Finally, the images collected by high-definition camera on fabric inspection machine were segmented, and the segmented small images were sent to the trained classification model in batches to realize the classification of each small image. Thereby the defects were detected and their positions were determined. The model was validated on a fabric defect database. The experimental results show that the average test time of each small picture is 0.37 ms by this proposed model, which is 67% lower than that by GoogLeNet, 93% lower than that by ResNet-50, and the accuracy of the proposed model is 99.99% on test set, which shows that its accuracy and real-time performance meet actual industrial demands.

Aiming at the inaccurate mining results of the Maximum Frequent Itemset mining algorithm based on Hopfield Neural Network (HNNMFI), an improved association rule mining algorithm for Hopfield neural network based on current ThrEshold Adaptive Memristor (TEAM) model was proposed. Firstly, TEAM model was used to design and implement synapses whose weights were set and updated by the ability of that threshold memristor continuously changes memristance value with square-wave voltage, and the input of association rule mining algorithm was self-adapted by the neural network. Secondly, the energy function was improved to align with standard energy function, and the memristance values were used to represent the weights, then the weights and bias were amplified. Finally, an algorithm of generating association rules from the maximum frequent itemsets was designed. A total of 1000 simulation experiments using 10 random transaction sets with size less than 30 were performed. Experimental results show that compared with HNNMFI algorithm, the proposed algorithm improves the result accuracy of association mining by more than 33.9%, which indicates that the memristor can effectively improve the result accuracy of Hopfield neural network in association rule mining.

A two-stage path planning algorithm was proposed for multi-Agent path planning. Firstly, an improved ant colony algorithm was used to plan an optimal path for each Agent from the starting point to the target point without colliding with the static obstacles in the environment. The reverse learning method was introduced to an improved ant colony algorithm to initialize the ant positions and increase the global search ability of the algorithm. The adaptive inertia weighted factor in the particle swarm optimization algorithm was used to adjust the pheromone intensity Q value to make it adaptively change to avoid falling into local optimum. The pheromone volatilization factor ρ was adjusted to speed up the iteration of the algorithm. Then, if there were dynamic collisions between multiple Agents, the game theory was used to construct a dynamic obstacle avoidance model between them, and the virtual action method was used to solve the game and select multiple Nash equilibria, making each Agent quickly learn the optimal Nash equilibrium. The simulation results show that the improved ant colony algorithm has a significant improvement in search accuracy and search speed compared with the traditional ant colony algorithm. And compared with Mylvaganam's multi-Agent dynamic obstacle avoidance algorithm, the proposed algorithm reduces the total path length and improves the convergence speed.

The problem of misclassification of minority class samples appears frequently when classifying massive amount of imbalanced data in real life with traditional classification algorithms, because most of these algorithms only suit balanced class distribution or samples with same misclassification cost. To overcome this problem, a classification algorithm for imbalanced dataset based on cost sensitive ensemble learning and oversampling-New Imbalanced Boost (NIBoost) was proposed. Firstly, the oversampling algorithm was used to add a certain number of minority samples to balance the dataset in each iteration, and the classifier was trained on the new dataset. Secondly, the classifier was used to classify the dataset to obtain the predicted class label of each sample and the classification error rate of the classifier. Finally, the weight coefficient of the classifier and new weight of each sample were calculated according to the classification error rate and the predicted class labeles. Experimental results on UCI datasets with decision tree and Naive Bayesian used as weak classifier algorithm show that when decision tree was used as the base classifier of NIBoost, compared with RareBoost algorithm, the F-value is increased up to 5.91 percentage points, the G-mean is increased up to 7.44 percentage points, and the AUC is increased up to 4.38 percentage points. The experimental results show that the proposed algorithm has advantages on imbalanced data classification problem.

When noise and intensity distortion exist in brain images, the method based on structural information cannot accurately extract image intensity information, edge and texture features at the same time. In addition, the computational complexity of continuous optimization is relatively high. To solve these problems, according to the structural information of the image, a non-rigid multi-modal brain image registration method based on Improved Zernike Moment based Local Descriptor (IZMLD) and Graph Cuts (GC) discrete optimization was proposed. Firstly, the image registration problem was regarded as the discrete label problem of Markov Random Field (MRF), and the energy function was constructed. The two energy terms were composed of the pixel similarity and smoothness of the displacement vector field. Secondly, a smoothness constraint based on the first derivative of the deformation vector field was used to penalize displacement labels with sharp changes between adjacent pixels. The similarity metric based on IZMLD was used as a data item to represent pixel similarity. Thirdly, the Zernike moments of the image patches were used to calculate the self-similarity of the reference image and the floating image in the local neighborhood and construct an effective local descriptor. The Sum of Absolute Difference (SAD) between the descriptors was taken as the similarity metric. Finally, the whole energy function was discretized and its minimum value was obtained by using an extended optimization algorithm of GC. The experimental results show that compared with the registration method based on the Sum of Squared Differences on Entropy images (ESSD), the Modality Independent Neighborhood Descriptor (MIND) and the Stochastic Second-Order Entropy Image (SSOEI), the mean of the target registration error of the proposed method was decreased by 18.78%, 10.26% and 8.89% respectively; and the registration time of the proposed method was shortened by about 20 s compared to the continuous optimization algorithm. The proposed method achieves efficient and accurate registration for images with noise and intensity distortion.

The ship behavior recognition by human supervision in complex marine environment is inefficient. In order to solve the problem, a new ship behavior recognition method based on multi-scale convolutional neural network was proposed. Firstly, massive ship driving data were obtained from the Automatic Identification System (AIS), and the discriminative ship behavior trajectories were extracted. Secondly, according to the characteristics of the trajectory data, the behavior recognition network for ship trajectory data was designed and implemented by multi-scale convolution, and the feature channel weighting and Long Short-Term Memory network (LSTM) were used to improve the accuracy of algorithm. The experimental results on ship behavior dataset show that, the proposed recognition network can achieve 92.1% recognition accuracy for the ship trajectories with specific length, which is 5.9 percentage points higher than that of the traditional convolutional neural network. In addition, the stability and convergence speed of the proposed network are significantly improved. The proposed method can effectively improve the ship behavior recognition accuracy, and provide efficient technical support for the marine regulatory authority.

In the traditional medical image fusion, it is necessary to manually set the fusion rules and parameters by using prior knowledge, which leads to the uncertainty of fusion effect and the lack of detail expression. In order to solve the problems, a Computed Tomography (CT)/Magnetic Resonance (MR) image fusion algorithm based on improved Generative Adversarial Network (GAN) was proposed. Firstly, the network structures of generator and discriminator were improved. In the design of generator network, residual block and fast connection were used to deepen the network structure, so as to better capture the deep image information. Then, the down-sampling layer of the traditional network was removed to reduce the information loss during image transmission, and the batch normalization was changed to the layer normalization to better retain the source image information, and the depth of the discriminator network was increased to improve the network performance. Finally, the CT image and the MR image were connected and input into the generator network to obtain the fused image, and the network parameters were continuously optimized through the loss function, and the model most suitable for medical image fusion was trained to generate the high-quality image. The experimental results show that, the proposed algorithm is superior to Discrete Wavelet Transformation (DWT) algorithm, NonSubsampled Contourlet Transform (NSCT) algorithm, Sparse Representation (SR) algorithm and Sparse Representation of classified image Patches (PSR) algorithm on Mutual Information (MI), Information Entropy (IE) and Structural SIMilarity (SSIM). The final fused image has rich texture and details. At the same time, the influence of human factors on the stability of the fusion effect is avoided.

In the traditional Convolutional Neural Network (CNN), the information cannot be transmitted to each other between the neurons of the same layer, the feature information at the same layer cannot be fully utilized, making the lack of the representation of the characteristics of the sentence system. As the result, the feature learning ability of model is limited and the text classification effect is influenced. Aiming at the problem, a model based on joint network CNN-BiGRU and attention mechanism was proposed. In the model, the CNN-BiGRU joint network was used for feature learning. Firstly, deep-level phrase features were extracted by CNN. Then, the Bidirectional Gated Recurrent Unit (BiGRU) was used for the serialized information learning to obtain the characteristics of the sentence system and strengthen the association of CNN pooling layer features. Finally, the effective feature filtering was completed by adding attention mechanism to the hidden state weighted calculation. Comparative experiments show that the method achieves 91.93% F1 value and effectively improves the accuracy of text classification with small time cost and good application ability.

Deep neural networks tend to suffer from overfitting problem because of the high complexity of the model. To reduce the adverse eeffects of the problem on the network performance, an improved elastic network model based deep learning optimization method was proposed. Firstly, considering the strong correlation between the variables, the adaptive weights were assigned to different variables of L1-norm in elastic network model, so that the linear combination of the L2-norm and the adaptively weighted L1-norm was obtained. Then, the solving process of neural network parameters under this new regularization term was given by combining improved elastic network model with the deep learning optimization model. Moreover, the robustness of this proposed model was theoretically demonstrated by showing the grouping selection ability and Oracle property of the improved elastic network model in the optimization of neural network. At last, in regression and classification experiments, the proposed model was compared with L1-norm, L2-norm and elastic network regularization term, and had the regression error decreased by 87.09, 88.54 and 47.02 and the classification accuracy improved by 3.98, 2.92 and 3.58 percentage points respectively. Thus, theory and experimental results prove that the improved elastic network model can effectively improve the generalization ability of deep neural network model and the performance of optimization algorithm, and solve the overfitting problem of deep learning.

Aiming at the ambiguity caused by the loss of details in the super-resolution reconstruction of medical images, a medical image super-resolution algorithm based on deep residual Generative Adversarial Network (GAN) was proposed. Firstly, a generative network and a discriminative network were designed in the method. High resolution images were generated by the generative network and the authenticities of the images were identified by the discriminative network. Secondly, a resize-convolution was used to eliminate checkerboard artifacts in the upsampling layer of the designed generative network and the batch-normalization layer of the standard residual block was removed to optimize the network. Also, the number of feature maps was further increased in the discriminative network and the network was deepened to improve the network performance. Finally, the network was continuously optimized according to the generative loss and the discriminative loss to guide the generation of high-quality images. The experimental results show that compared with bilinear interpolation, nearest-neighbor interpolation, bicubic interpolation, deeply-recursive convolutional network for image super-resolution and Super-Resolution using a Generative Adversarial Network (SRGAN), the improved algorithm can reconstruct the images with richer texture and more realistic vision. Compared with SRGAN, the proposed algorithm has an increase of 0.21 dB and 0.32% in Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM). It provides a deep residual generative adversarial network method for the theoretical research of medical image super-resolution, which is reliable and effective in practical applications.

When using the in-network nodes with cache ability for collaborative caching, the packets need to be forward quickly according to the surrounding caching status. A new data-plane-fast-forwarding method was proposed for this problem. Two bloom filters were kept for each port in the switch to maintain the surrounding caching status at the data plane. Meanwhile, the action of protocol oblivious forwarding was also extended. The extended action searched the bloom filters directly, and the optimized forwarding process was used to forward packets according to the searching results, then the packets were forwarded quickly based on the surrounding caching status. The evaluation results show that the caching status maintained by the controller reaches the forwarding performance bottleneck when the input rate is 80 Kb/s. The packets can be forwarded at line speed when the input rate is 111 Mb/s by using the data-plane-fast-forwarding method, which efficiency of forwarding is superior to the output action of protocol oblivious forwarding. The memory overhead of maintaining caching status by using the bloom filter is up to 20% of that by using the flow table. In Software Defined Networking (SDN) with cache ability, the proposed method can maintain the surrounding caching status at the data plane and promote the efficiency of forwarding packets by the surrounding caching status for collaborative caching.

The process of algal bloom evolution in rivers or lakes has characteristics of suddenness and uncertainty, which leads to low prediction accuracy of algal bloom. To solve this problem, chlorophyll a concentration was used as the surface index of cyanobacteria bloom evolution process, and a cyanobacterial bloom forecast model based on Long Short-Term Memory (LSTM) and Recurrent Neural Network (RNN) was proposed. Firstly, the improved Genetic algorithm-First order lag filter (GF) optimization algorithm was taken as data smoothing filter. Secondly, a GF-LSTM network model was built to accurately predict the cyanobacterial bloom. Finally, the data sampled from Meiliang Lake in Taihu area were used to test the forecast model, and then the model was compared with the traditional RNN and LSTM network. The experimental results show that, the mean relative error of the proposed GF-LSTM network model is 16%-18%, lower than those of RNN model (28%-32%) and LSTM network model (19%-22%). The proposed model has good effect on data smoothing filtering, higher prediction accuracy and better adaptability to samples. It also avoids two widely known issues of gradient vanishing and gradient exploding when using traditional RNN model during long term training.

In order to solve the problem of updating distance information frequently in the traditional method of setting uplink synchronization based on ranging information, a uplink clock synchronization method based on location information was proposed. Firstly, by measuring the pseudoranges to form a nonlinear system of equations, the location information of the terrestrial unit was located by using the solution method based on the principle of least squares. Then, due to the known location information of satellite movement, the change relationship of the distance between the satellite and the ground with time could be further obtained. The distance was converted into time delay to obtain the time advance of the uplink signal transmission of the terrestrial unit. Finally, the transmitter of the terrestrial unit was adjusted to make that the uplink signal could just arrive at the satellite in the assigned time slot with high accuracy, and the purpose of uplink clock synchronization was achieved. The simulation results show that, the proposed method can realize the synchronization of uplink clock in the satellite constellation communication system with high accuracy for the static units in the earth surface all over the world, and avoid the frequent ranging updates with high accuracy.

Industrial waste gas has accounted for about 70% of the atmospheric pollution sources. It is crucial to establish a full-scale and reasonable monitoring mechanism. However, the monitoring area is so large and monitoring devices can not be set up in some special areas. Besides, it is difficult to model the gas distribution according with the actual. To solve the practical and theoretical problems, an analysis method of industrial waste gas distribution for non-detection zone was proposed based on a Bi-directional Error Multi-Layer Neural Network (BEMNN). Firstly, the monitoring mechanism was introduced in the thought of "monitoring in boundary and inference of dead zone", which aimed to offset the lack of monitoring points in some areas. Secondly, a multi-layer combination neural network was proposed in which the errors propagate in a bi-directional mode. The network was used to model the gas distribution relationship between the boundary and the dead zone. Then the gas distribution in the dead zone could be predicted with the boundary monitoring data. Finally, an experiment was conducted based on the actual monitoring data of an industrial park. The mean absolute error was less than 28.83 μg and the root-mean-square error was less than 45.62 μg. The relative error was between 8% and 8.88%. The results prove the feasibility of the proposed method, which accuracy can meet the practical requirement.