A multi-scale 2D-Adaboost algorithm was proposed to solve the problem that the microscopic images of Chinese medicinal materials powder contain a large number of fine features and background interference factors， which leads to excessive changes in the same medicinal materials （large differences within the class） and too similar features among various medicinal materials （small differences between the classes）. Firstly， a global-local feature fusion backbone network architecture was constructed to extract multi-scale features better. By combining the advantages of Transformer and Convolutional Neural Network （CNN）， this architecture was able to extract and fuse global and local features at various scales effectively， thereby improving the feature capture capability of the backbone network significantly. Secondly， the single-scale output of Adaboost was extended to multi-scale output， and a 2D-Adaboost structure-based background suppression module was constructed. With this module， the output feature maps of each scale of the backbone network were divided into foreground and background， thereby suppressing feature values of the background region effectively and enhancing the strength of discriminative features. Finally， an extra classifier was added to each scale of the 2D-Adaboost structure to build a feature refinement module， which coordinated the collaborative learning among the classifiers by controlling temperature parameters， thereby refining the feature maps of different scales gradually， helping the network to learn more appropriate feature scales， and enriching the detailed feature representation. Experimental results show that the recognition accuracy of the proposed algorithm reaches 96.85%， which is increased by 7.56， 5.26， 3.79 and 2.60 percentage points， respectively， compared with those of ConvNeXt-L， ViT-L， Swin-L， and Conformer-L models. The high accuracy and stability of the classification validate the effectiveness of the proposed algorithm in classification tasks of Chinese medicinal materials powder microscopic images.

In today’s society with the growing demand for privacy protection， federated learning is receiving widespread attention. However， in federated learning， it is difficult for the server to supervise behaviors of clients， so that the existence of lazy clients poses a potential threat to the performance and fairness of federated learning. Aiming at the problem of how to identify lazy clients efficiently and accurately， a dual-task proof-of-work method based on backdoor was proposed， namely FedBD （FedBackDoor）. In FedBD， additional backdoor tasks that are easier to detect were allocated by the server for the clients participating in federated learning， the backdoor tasks were trained by the clients based on the original training tasks， and the clients’ behaviors were supervised by the server indirectly through training status of the backdoor tasks. Experimental results show that FedBD has certain advantages over the classic federated averaging algorithm FedAvg and the advanced algorithm GTG-Shapley （Guided Truncation Gradient Shapley） on datasets such as MNIST and CIFAR10. On CIFAR10 dataset， when the proportion of lazy clients is 15%， FedBD improves the accuracy by more than 10 percentage points compared with FedAvg， and increases the accuracy by 2 percentage points compared with GTG-Shapley. Moreover， the average training time of FedBD is only 11.8% of that of GTG-Shapley， and the accuracy of FedBD in identifying lazy clients can exceed 99% when the proportion of lazy clients is 10%. It can be seen that FedBD can solve the problem of lazy clients being difficult to supervise.

The diagnosis of major depressive disorder predominantly relies on subjective methods， including physician consultations and scale assessments， which may lead to misdiagnosis. EEG （ElectroEncephaloGraphy） offers advantages such as high temporal resolution， low cost， ease of setup， and non-invasiveness， making it a potential quantitative measurement tool for psychiatric disorders， including depressive disorder. Recently， deep learning algorithms have been diversely applied to EEG signals， notably in the diagnosis and classification of depressive disorder. Due to significant redundancy is observed when processing EEG signals through a self-attention mechanism， a convolutional neural network leveraging a Probabilistic sparse Self-Attention mechanism （PSANet） was proposed. Firstly， a limited number of pivotal attention points were chosen in the self-attention mechanism based on the sampling factor， addressing the high computational cost and facilitating its application to extensive EEG data sequences； concurrently， EEG data was amalgamated with patients’ physiological scales for a comprehensive diagnosis. Experiments were executed on a dataset encompassing both depressive disorder patients and a healthy control group. Experimental results show that PSANet exhibits superior classification accuracy and a reduced number of parameters relative to alternative methodologies such as EEGNet.

Aiming to the problems of low embedding capacity and poor visual quality of the extracted secret images in existing generative data hiding algorithms， a generative data hiding algorithm based on multi-scale attention was proposed. First， a generator with dual encode-single decode based on multi-scale attention was designed. The features of the cover image and secret image were extracted independently at the encoding end in two branches， and fused at the decoding end by a multi-scale attention module. Skip connections were used to provide different scales of detail features， thereby ensuring high-quality of the stego-image. Second， self-attention module was introduced into the extractor of the U-Net structure to weaken the deep features of the cover image and enhance the deep features of the secret image. The skip connections were used to compensate for the detail features of the secret image， so as to improve the accuracy of the extracted secret data. At the same time， the adversarial training of the multi-scale discriminator and generator could effectively improve the visual quality of the stego-image. Experimental results show that the proposed algorithm can achieve an average Peak Signal-to-Noise Ratio （PSNR） and Structure Similarity Index Measure （SSIM） of 40.93 dB and 0.988 3 for the generated stego-images， and an average PSNR and SSIM of 30.47 dB and 0.954 3 for the extracted secret images under the embedding capacity of 24 bpp.

Existing emotion recognition models based on Electroencephalogram （EEG） almost ignore differences in emotional states at different time periods， and fail to reinforce key emotional information. To solve the above problem， a Multiple Context Vectors optimized Convolutional Recurrent neural network （CR-MCV）was proposed. Firstly， feature matrix sequence of EEG signals was constructed to obtain spatial features of multi-channel EEG by Convolutional Neural Network （CNN）. Then， recurrent neural network based on multi-head attention was adopted to generate multiple context vectors for high-level abstract feature extraction. Finally， a fully connected layer was used for emotion classification. Experiments were carried out on DEAP （Database for Emotion Analysis using Physiological signals） dataset， and the classification accuracy in arousal and valence dimensions was 88.09% and 89.30%， respectively. Experimental results show that the CR-MCV can adaptively allocate attention of features and strengthen salient information related to emotion states based on utilization of electrode spatial position information and saliency characteristics of emotional states at different time periods.

A real-time object detection algorithm YOLO-C for complex construction environment was proposed for the problems of cluttered environment， obscured objects， large object scale range， unbalanced positive and negative samples， and insufficient real-time of existing detection algorithms， which commonly exist in construction environment. The extracted low-level features were fused with the high-level features to enhance the global sensing capability of the network， and a small object detection layer was designed to improve the detection accuracy of the algorithm for objects of different scales. A Channel-Spatial Attention （CSA） module was designed to enhance the object features and suppress the background features. In the loss function part， VariFocal Loss was used to calculate the classification loss to solve the problem of positive and negative sample imbalance. GhostConv was used as the basic convolutional block to construct the GCSP （Ghost Cross Stage Partial） structure to reduce the number of parameters and the amount of computation. For complex construction environments， a concrete construction site object detection dataset was constructed， and comparison experiments for various algorithms were conducted on the constructed dataset. Experimental results demonstrate that the YOLO?C has higher detection accuracy and smaller parameters， making it more suitable for object detection tasks in complex construction environments.

Enhancing the robustness of complex networks is crucial for the networks to resist external attacks and cascading failures. Existing evolutionary algorithms have limitations in solving network structure optimization problems， especially in terms of convergence and optimization speed. In response to this challenge， a new adaptive complex network robustness optimization algorithm named SU-ANet （SUrrogate-assisted and Adaptive Network optimization algorithm） was proposed. To reduce the huge time overhead of robustness computation， a robustness predictor based on attention mechanism was constructed in SU-ANet as an offline surrogate model to replace the frequent robustness computation in local search operator. In the evolutionary process， the global and local information was considered comprehensively to avoid falling into local optimum and broaden the search space simultaneously. By designing crossover operators， each individual exchanges edges with the global optimum candidate solution and a randomly selected individual to balance the convergence and diversity of the algorithm. Additionally， a parameter self-adaptive mechanism was applied to adjust the operator execution probabilities automatically， thereby alleviating the uncertainty of the algorithm brought by the parameter design. Experimental results on both synthetic networks and real-world networks demonstrate that SU-ANet has better search capabilities and higher evolutionary efficiency.

To address the problem of difficulty in identifying user plagiarism in social networks and to protect the rights of original authors while holding users accountable for plagiarism actions， a plagiarism identification scheme for social network users under blockchain was proposed. Aiming at the lack of universal tracing model in existing blockchain， a blockchain-based traceability information management model was designed to record user operation information and provide a basis for text similarity detection. Based on the Merkle tree and Bloom filter structures， a new index structure BHMerkle was designed. The calculation overhead of block construction and query was reduced， and the rapid positioning of transactions was realized. At the same time， a multi-feature weighted Simhash algorithm was proposed to improve the precision of word weight calculation and the efficiency of signature value matching stage. In this way， malicious users with plagiarism cloud be identified， and the occurrence of malicious behavior can be curbed through the reward and punishment mechanism. The average precision and recall of the plagiarism detection scheme on news datasets with different topics were 94.8% and 88.3%， respectively. Compared with multi-dimensional Simhash algorithm and Simhash algorithm based on information Entropy weighting （E-Simhash）， the average precision was increased by 6.19 and 4.01 percentage points respectively， the average recall was increased by 3.12 and 2.92 percentage points respectively. Experimental results show that the proposed scheme improves the query and detection efficiency of plagiarism text， and has high accuracy in plagiarism identification.

The complexity of pedestrian interaction is a challenge for pedestrian trajectory prediction， and the existing algorithms are difficult to capture meaningful interaction information between pedestrians， which cannot intuitively model the interaction between pedestrians. To address this problem， a multi-head soft attention graph convolutional network was proposed. Firstly， a Multi-head Soft ATTention （MS ATT） combined with involution network was used to extract sparse spatial adjacency matrix and sparse temporal adjacency matrix from spatial and temporal graph inputs respectively to generate sparse spatial directed graph and sparse temporal directed graph. Then， a Graph Convolutional Network （GCN） was used to learn interaction and motion trend features from sparse spatial and sparse temporal directed graphs. Finally， the learned trajectory features were input into a Temporal Convolutional Network （TCN） to predict double Gaussian distribution parameters， thereby generating the predicted pedestrian trajectories. Experiments on Eidgenossische Technische Hochschule （ETH） and University of CYprus （UCY） datasets show that， compared with Space-time sOcial relationship pooling pedestrian trajectory Prediction Model （SOPM）， the proposed algorithm reduces the Average Displacement Error （ADE） by 2.78%， and compared to Sparse Graph Convolution Network （SGCN）， the proposed algorithm reduces the Final Displacement Error （FDE） by 16.92%.

Sparse-dense Matrix Multiplication （SpMM） is widely used in the fields such as scientific computing and deep learning， and it is of great importance to improve its efficiency. For a class of sparse matrices with band feature， a new storage format BRCV （Banded Row Column Value） and an SpMM algorithm based on this format as well as an efficient Graphics Processing Unit （GPU） implementation were proposed. Due to the fact that each sparse band can contain multiple sparse blocks， the proposed format can be seen as a generalization of the block sparse matrix format. Compared with the commonly used CSR （Compressed Sparse Row） format， BRCV format was able to significantly reduce the storage complexity by avoiding redundant storage of column indices in sparse bands. At the same time， the GPU implementation of SpMM based on BRCV format was able to make more efficient use of GPU’s shared memory and improve the computational efficiency of SpMM algorithm by reusing the rows of both sparse and dense matrices. For randomly generated band sparse matrices， experimental results on two different GPU platforms show that BRCV outperforms not only cuBLAS （CUDA Basic Linear Algebra Subroutines）， but also cuSPARSE based on CSR and block sparse formats. Specifically， compared with cuSPARSE based on CSR format， BRCV has the maximum speedup ratio of 6.20 and 4.77 respectively. Moreover， the new implementation was applied to accelerate the SpMM operator in Graph Neural Network （GNN）. Experimental results on real application datasets show that BRCV outperforms cuBLAS and cuSPARSE based on CSR format， also outperforms cuSPARSE based on block sparse format in most cases. In specific， compared with cuSPARSE based on CSR format， BRCV has the maximum speedup ratio reached 4.47. The above results indicate that BRCV can improve the efficiency of SpMM effectively.

The existing similarity-based moving target trajectory prediction algorithms are generally classified according to the spatial-temporal characteristics of the data， and the characteristics of the algorithms themselves cannot be reflected. Therefore， a classification method based on algorithm characteristics was proposed. The calculation of the distances between two points is required for the trajectory similarity algorithms to carry out the subsequent calculations， however， the commonly used Euclidean Distance （ED） is only applicable to the problem of moving targets in a small region. A method of similarity calculation using geodetic distance instead of ED was proposed for the trajectory prediction of sea targets moving in a large region. Firstly， the trajectory data were preprocessed and segmented. Then， the discrete Fréchet Distance （FD） was adopted as similarity measure. Finally， synthetic and real data were used to test. Experimental results indicate that when sea targets move in a large region， the ED-based algorithm may gain incorrect prediction results， while the geodetic distance-based algorithm can output correct trajectory prediction.

Attackers can illegally open a vehicle by forgeing the Radio Frequency IDentification （RFID） signal sent by the vehicle remote key. Besides， when the vehicle remote key is lost or stolen， the attacker can obtain the secret data inside the vehicle remote key and clone a usable vehicle remote key， which will threaten the property and privacy security of the vehicle owner. Aiming at the above problems， a Vehicle RKE Two-Factor Authentication （VRTFA） protocol for vehicle Remote Keyless Entry （RKE） that resists physical cloning attack was proposed. The protocol is based on Physical Uncloneable Function （PUF） and biological fingerprint feature extraction and recovery functions， so that the specific hardware physical structure of the legal vehicle remote key cannot be forged. At the same time， the biological fingerprint factor was introduced to build a two-factor authentication protocol， thereby solving the security risk of vehicle remote key theft， and further guaranteeing the secure mutual authentication of vehicle RKE system. Security analysis results of the protocol using BAN logic show that VRTFA protocol can resist malicious attacks such as forgery attack， desynchronization attack， replay attack， man-in-the-middle attack， physical cloning attack， and full key leakage attack， and satisfy the security attributes such as forward security， mutual authentication， data integrity， and untraceability. Performance analysis results show that VRTFA protocol has stronger security and privacy and better practicality than the existing RFID authentication protocols.

Ring signature is widely used to solve the problems of user identity and data privacy disclosure because of its spontaneity and anonymity； and certificateless public key cryptosystem can not only solve the problem of key escrow， but also do not need the management of public key certificates； certificateless ring signature combines the advantages of both of the above mentioned， and has extensive research significance， but most of the existing certificateless ring signature schemes are based on the calculation of bilinear pairings and modular exponentiation， which are computationally expensive and inefficient. In order to improve the efficiency of signature and verification stages， a new Efficient CertificateLess Ring Signature （ECL-RS） scheme was proposed， which used elliptic curve with low computational cost， high security and good flexibility. The security statute of ECL-RS scheme stems from a discrete logarithm problem and a Diffie-Hellman problem， and the scheme is proved to be resistant to public key substitution attacks and malicious key generation center attacks under Random Oracle Model （ROM） with unforgeability and anonymity. Performance analysis shows that ECL-RS scheme only needs （n+2） （n is the number of ring members） elliptic curve scalar multiplication and scalar addition operations as well as （n+3） one-way hash operations， which has lower computational cost and higher efficiency while ensuring security.

Cross-project software defect prediction can solve the problem of few training data in prediction projects. However， the source project and the target project usually have the large distribution difference， which reduces the prediction performance. In order to solve the problem， a new Cross-Project Defect Prediction method based on Feature Selection and TrAdaBoost （CPDP-FSTr） was proposed. Firstly， in the feature selection stage， Kernel Principal Component Analysis （KPCA） was used to delete redundant data in the source project. Then， according to the attribute feature distribution of the source project and the target project， the candidate source project data closest to the target project distribution were selected according to the distance. Finally， in the instance transfer stage， the TrAdaBoost method improved by the evaluation factor was used to find out the instances in the source project which were similar to the distribution of a few labeled instances in the target project， and establish a defect prediction model. Using F1 as the evaluation index， compared with the methods such as cross-project software defect prediction using Feature Clustering and TrAdaBoost （FeCTrA）， Cross-project software defect prediction based on Multiple Kernel Ensemble Learning （CMKEL）， the proposed CPDP-FSTr had the prediction performance improved by 5.84% and 105.42% respectively on AEEEM dataset， enhanced by 5.25% and 85.97% respectively on NASA dataset， and its two-process feature selection is better than the single feature selection process. Experimental results show that the proposed CPDP-FSTr can achieve better prediction performance when the source project feature selection proportion and the target project labeled instance proportion are 60% and 20% respectively.

Most of the existing network embedding methods only preserve the local structure information of the network， while they ignore other potential information in the network. In order to preserve the community information of the network and reflect the multi-granularity characteristics of the network community structure， a network Embedding method based on Multi-Granularity Community information （EMGC） was proposed. Firstly， the network’s multi-granularity community structure was obtained， the node embedding and the community embedding were initialized. Then， according to the node embedding at previous level of granularity and the community structure at this level of granularity， the community embedding was updated， and the corresponding node embedding was adjusted. Finally， the node embeddings under different community granularities were spliced to obtain the network embedding that fused the community information of different granularities. Experiments on four real network datasets were carried out. Compared with the methods that do not consider community information （DeepWalk， node2vec） and the methods that consider single-granularity community information （ComE， GEMSEC）， EMGC’s AUC value on link prediction and F1 score on node classification are generally better than those of the comparison methods. The experimental results show that EMGC can effectively improve the accuracy of subsequent link prediction and node classification.

Aiming at the problem of model information leakage caused by interpretability in Deep Neural Network （DNN）， the feasibility of using the Gradient-weighted Class Activation Mapping （Grad-CAM） interpretation method to generate adversarial samples in a white-box environment was proved， moreover， an untargeted black-box attack algorithm named dynamic genetic algorithm was proposed. In the algorithm， first， the fitness function was improved according to the changing relationship between the interpretation area and the positions of the disturbed pixels. Then， through multiple rounds of genetic algorithm， the disturbance value was continuously reduced while increasing the number of the disturbed pixels， and the set of result coordinates of each round would be maintained and used in the next round of iteration until the perturbed pixel set caused the predicted label to be flipped without exceeding the perturbation boundary. In the experiment part， the average attack success rate under the AlexNet， VGG-19， ResNet-50 and SqueezeNet models of the proposed algorithm was 92.88%， which was increased by 16.53 percentage points compared with that of One pixel algorithm， although with the running time increased by 8% compared with that of One pixel algorithm. In addition， in a shorter running time， the proposed algorithm had the success rate higher than the Adaptive Fast Gradient Sign Method （Ada-FGSM） algorithm by 3.18 percentage points， higher than the Projection & Probability-driven Black-box Attack （PPBA） algorithm by 8.63 percentage points， and not much different from Boundary-attack algorithm. The results show that the dynamic genetic algorithm based on the interpretation method can effectively execute the adversarial attack.

In order to solve the problems that the iterative efficiencies of the existing privacy protection k-means clustering schemes are low， the server in the centralized differential privacy preserving k-means clustering scheme may be attacked， and the server in the localized differential privacy protection k-means clustering scheme may return wrong clustering results， a Multi-party Privacy Protection k-means Clustering Scheme based on Blockchain （M-PPkCS/B） was proposed. Taking advantages of localized differential privacy technology and the characteristics of the blockchain such as being open， transparent， and non-tamperable， firstly， a Multi-party k-means Clustering Center Initialization Algorithm （M-kCCIA） was designed to improve the iterative efficiency of clustering while protecting user privacy， and ensure the correctness of initial clustering centers jointly generated by the users. Then， a Blockchain-based Privacy Protection k-means Clustering Algorithm （Bc-PPkCA） was designed， and a smart contract of clustering center updating algorithm was constructed. The clustering center was updated iteratively by the above smart contract on the blockchain to ensure that each user was able to obtain the correct clustering results. Through experiments on the datasets HTRU2 and Abalone， the results show that while ensuring that each user obtains the correct clustering results， the accuracy can reach 97.53% and 96.19% respectively， the average iteration times of M-kCCIA is 5.68 times and 2.75 times less than that of the algorithm of randomly generating initial cluster center called Random Selection （RS）.

Aiming at the limited scalability of medical data sharing based on traditional blockchains， a scale-out and sharing scheme of blockchain based on sharding technology was proposed. Firstly， the periodic network sharding was performed based on the jump consistent hash algorithm， and the risk of Sybil attacks in a single shard was greatly reduced by randomly dividing the network nodes. Then， the Scalable decentralized Trust inFrastructure for Blockchains （SBFT） consensus protocol was used in the shards to reduce the high communication complexity of the Pratic Byzantic Fault Torent （PBFT） consensus protocol， and the two-layer architecture was used between the physical multi-chain of shards and the logical single chain of the main chain to reduce the storage pressure of the members of shards. Finally， a multi-keyword association retrieval searchable encryption sharing scheme based on Public key Encryption with Conjunctive field Keyword Search （PECKS） was proposed on the medical consortium blockchain， so as to improve the patients’ control over their sensitive data， and realize the fine-grained search of sensitive data under encryption. Through performance analysis， it can be seen that under the parallel sharding structure， the throughput of blockchain is significantly increased with the increase of shards， and the retrieval efficiency is also significantly improved. Experimental results show that the proposed scheme can greatly improve the efficiency and scalability of the blockchain system.

In the method of combining Object detection Network (ObjectNet) and scene recognition network, the object features extracted by the ObjectNet and the scene features extracted by the scene network are inconsistent in dimensionality and property, and there is redundant information in the object features that affects the scene judgment, resulting in low recognition accuracy of scenes. To solve this problem, an improved indoor scene recognition method combined with object detection was proposed. First, the Class Conversion Matrix (CCM) was introduced into the ObjectNet to convert the object features output by ObjectNet, so that the dimension of the object features was consistent with that of the scene features, as a result, the information loss caused by inconsistency of the feature dimensions was reduced. Then, the Context Gating (CG) mechanism was used to suppress the redundant information in the features, reducing the weight of irrelevant information, and increasing the contribution of object features in scene recognition. The recognition accuracy of the proposed method on MIT Indoor67 dataset reaches 90.28%, which is 0.77 percentage points higher than that of Spatial-layout-maintained Object Semantics Features (SOSF) method; and the recognition accuracy of the proposed method on SUN397 dataset is 81.15%, which is 1.49 percentage points higher than that of Hierarchy of Alternating Specialists (HoAS) method. Experimental results show that the proposed method improves the accuracy of indoor scene recognition.

Thread-Level Speculation (TLS) technology can tap the parallel execution potential of programs and improve the utilization of multi-core resources. However, the current TACLeBench kernel benchmarks are not effectively analyzed in TLS parallelization. In response to this problem, the loop-level speculative execution analysis scheme and analysis tool were designed. With 7 representative TACLeBench kernel benchmarks selected, firstly, the initialization analysis was performed to the programs, the program hot fragments were selected to insert the loop identifier. Then, the cross-compilation was performed to these fragments, the program speculative thread and the memory address related data were recorded, and the maximun potential of the loop-level parallelism was analyzed. Finally, the program runtime characteristics (thread granularity, parallelizable coverage, dependency characteristics) and the impacts of the source code on the speedup ratio were comprehensively discussed. Experimental results show that:1) this type of programs is suitable for TLS acceleration, compared with serial execution results, under the loop structure speculative execution, the speedup ratios for most programs are above 2, and the highest speedup ratio in them can reach 20.79; 2) by using TLS to accelerate the TACLeBench kernel programs, most applications can effectively make use of 4-core to 16-core computing resources.

In view of the problems of traditional file sharing schemes， such as easy leakage of files， difficult control of file destination， and complex access control， as well as the application requirements of cloud file hierarchical classification management and sharing， a hierarchical file access control scheme with identity-based multi-conditional proxy re-encryption was proposed. Firstly， the permission level of file was taken as the condition of ciphertext generation， and the trusted hierarchical management unit was introduced to determine and manage the user levels. Secondly， the re-encryption key of user’s hierarchical access permission was generated， which solved the problem that the identity-based conditional proxy re-encryption scheme only restricts the re-encryption behavior of proxy servers， and lacks the limitation of the user’s permission. Meanwhile， the burden of client was reduced， which means only encryption and decryption operations were needed for users. The results of comparison and analysis of different schemes show that， compared with the existing access control schemes， the proposed scheme has obvious advantages， it can complete the update of the user’s access permission without the direct participation of users， and has the characteristic of uploader anonymity.

In order to make full use of appearance and motion information in video abnormal behavior detection, a Siamese network model that can capture appearance and motion information at the same time was proposed. The two branches of the network were composed of the same autoencoder structure. Several consecutive frames of RGB images were used as the input of the appearance sub-network to predict the next frame, while RGB frame difference image was used as the input of the motion sub-network to predict the future frame difference. In addition, considering one of the reasons that affected the detection effect of the prediction-based method, that is the diversity of normal samples, and the powerful "generation" ability of the autoencoder network, that is it has a good prediction effect on some abnormal samples. Therefore, a memory enhancement module that learns and stores the "prototype" features of normal samples was added between the encoder and the decoder, so that the abnormal samples were able to obtain greater prediction error. Extensive experiments were conducted on three public anomaly detection datasets Avenue, UCSD-ped2 and ShanghaiTech. Experimental results show that, compared with other video abnormal behavior detection methods based on reconstruction or prediction, the proposed method achieves better performance. Specifically, the average Area Under Curve (AUC) of the proposed method on Avenue, UCSD-ped2 and ShanghaiTech datasets reach 88.2%, 97.5% and 73.0% respectively.

Sentiment analysis of massive text information on social networks can better mine the behavior rules of Internet users,helping decision-making institutions understand the public opinion tendencies and helping businesses improve the quality of service. The task of Chinese implicit sentiment classification is more difficult than those of other languages due to the absence of key emotional features,expression vector forms and cultural customs. The existing Chinese implicit sentiment classification methods are mainly based on Convolutional Neural Network(CNN),and have some defects, such as the inability to obtain the sequence of words and not using contextual emotional features reasonably in implicit emotion discrimination. A Chinese implicit sentiment classification model combining sequence and contextual features named GGBA (GCNN-GRU-BiGRU-Attention) was proposed to solve the above problems. In the model, Gated Convolutional Neural Network (GCNN) was used to extract the local important information of sentences with implicit sentiments,and Gated Recurrent Unit(GRU)network was used to enhance the temporal information of features. In the context feature processing of sentences with implicit sentiments,the combination of Bidirectional Gated Recurrent Unit (BiGRU)and attention was used to extract the important emotional features. After obtaining the two types of features,the contextual important features were integrated into the implicit emotion discrimination through the fusion layer. Experimental results on the implicit sentiment analysis evaluation dataset showed that the macro average precision of GGBA model was 3. 72% higher than that of normal text CNN named TextCNN,2. 57% higher than that of GRU,and 1. 90% higher than that of Disconnected Recurrent Neural Network(DRNN). Therefore,GGBA model achieves better classification performance than the basic models in implicit sentiment analysis tasks.

Concerning the problems that Teaching & Learning Based Optimization (TLBO) algorithm has slow convergence rate and low accuracy, and it is easy to be trapped into local optimum in solving high-dimensional problems, an Improved TLBO algorithm with Brain Storming Optimization (ITLBOBSO) was proposed. In this algorithm, a new “learning”operator was designed and applied to replace the origin “learning” in the TLBO. In the iteration process of the population, the “teaching” operator was executed by the current individual. Then, two individuals were selected randomly from the population, and brain storming learning was executed by the better one of the above and the current individual to improve the state of the current individual. Cauchy mutation and a random parameter associated with the iterations were introduced in the formula of this operator to improve the exploration ability in early stage and the exploitation ability for new solutions in later stage of the algorithm. In a series of simulation experimentations, compared with TLBO, the proposed algorithm has large improvements of solution accuracy, robustness and convergence speed on 11 benchmark functions. The experimental results on two constrained engineering optimization problems show that compared to TLBO algorithm, ITLBOBSO reduces the total cost by 4 percentage points, which proves the effectiveness of the proposed mechanism on overcoming the weakness of TLBO algorithm. The proposed algorithm is suitable for solving high dimensional continuous optimization problems.

Aiming at the problems of long registration time and low registration accuracy of current image registration algorithms,an image registration algorithm based on Grid-based Motion Statistics(GMS),Vector Coefficient Similarity (VCS)and Graph-Cut RANdom SAmple Consensus(GC-RANSAC)was proposed. Firstly,the feature points of the image were extracted through the ORB(Oriented FAST and Rotated BRIEF)algorithm,and Brute-Force matching of the feature points was performed. Then,the coarse matching feature points in the image were meshed by the GMS algorithm,and the coarse matching pairs were filtered based on the principle that high feature support exists in the neighborhood of the correct matching points in the grid. And the part elimination was performed to the matching pairs by introducing the principle that the image matching pair has VCS not exceed a set threshold during vector operation,which is beneficial to the fast convergence of the algorithm in the later stage. Finally,the local optimal model fitting was performed by using the GC-RANSAC algorithm to obtain the fine matching feature point set and achieve image registration and stitching with high precision. Compared with algorithms such as ASIFT+RANSAC,GMS,AKAZE+RANSAC,GMS+GC-RANSAC,the results show that the proposed algorithm improves the average matching accuracy by 30. 34% and reduces the average matching time by 0. 54 s.

Compressed sensing mainly contains random projection and reconstruction. Because of lower convergence speed of iterative shrinkage algorithm and the lacking of direction of traditional 2-dimensional wavelet transform， random projection was implemented by using Permute Discrete Cosine Transform （PDCT）， and the gradient projection was used for reconstruction. Based on the simplification of computation complexity， the transformation coefficients in the dual-tree complex wavelet domain were improved by iteration. Finally， the reconstructed image was obtained by the inverse transform. In the experiments， the reconstruction results of DT CWT （Dual-Tree Complex Wavelet Transform） and bi-orthogonal wavelet were compared with the same reconstruction algorithm， and the former is better than the latter in image detail and smoothness with higher Peak Signal-to-Noise Ratio （PSNR） of 1.5 dB. In the same sparse domain， gradient projection converges faster than iterative shrinkage algorithm. And in the same sparse domain and random projection， PDCT has a slightly higher PSNR than the structural random matrix.

For the monitoring dead angles of fixed surveillance cameras installed in large quantities and low hardware performance of mobile devices, an urban management case target recognition algorithm that can run on IOS mobile devices with low performance was proposed. Firstly, the number of channels of input and output images and the number of feature maps generated by each channel were optimized by adding new hyperparameters to MobileNet. Secondly, a new recognition algorithm was formed by combining the improved MobileNet with the SSD recognition framework and was transplanted to the IOS mobile devices. Finally, the accurate detection of the common 8 specific urban management case targets was achieved by the proposed algorithm, in which the camera provided by the mobile device was used to capture the scene video. The mean Average Precision (mAP) of the proposed algorithm was 15.5 percentage points and 10.4 percentage points higher than that of the prototype YOLO and the prototype SSD, respectively. Experimental results show that the proposed algorithm can run smoothly on low-performance IOS mobile devices, reduce the dead angles of monitoring, and provide technical support for urban management team to speed up the classification and processing of cases.

Focusing on the order allocation in vehicle logistics service supply chain, a bi-level programming model considering multiple modes of transportation was proposed. Firstly, considering that different transportation modes affect the transportation cost and the customer's on-time delivery requirement, a bi-level programming model aiming to punctual delivery and minimization of purchasing cost was established. Secondly, a Heuristic Algorithm (HA) was designed to determine the tasks of each transportation mode. Thirdly, Shuffled Frog Leaping Algorithm (SFLA) was used to solve task allocation of each transportation mode between functional logistics service providers. Finally, the solution of the proposed model was compared with those of Genetic Algorithm (GA), Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO) through different scale examples. The results show that compared with the original purchasing cost 4.38 million yuan, the proposed model has a significantly optimized result 4.21 million yuan, which shows the order allocation scheme of the proposed model solves the order allocation problem of vehicle logistics more effectively. Experimantal results show that HA-SFLA can obtain the significantly optimized result quickly compared to GA, PSO and ACO, illustrating that HA-SFLA can solve the bi-level model considering transportation modes more efficiently. The bi-level order allocation model and algorithm considering transportation modes can reduce logistics costs while meet customer on-time requirements, making the logistics suppliers consider the transportation modes in order allocation phase to achieve more benefits.

To efficiently detect outliers in massive hydrologic sensor data, an anomaly detection method for hydrological time series based on SparkR was proposed. Firstly, a sliding window and Autoregressive Integrated Moving Average (ARIMA) model were used to forecast the cleaned data on SparkR platform. Then, the confidence interval was calculated for the prediction results, and the results outside the interval range were judged as anomaly data. Finally, based on the detection results, K-Means algorithm was used to cluster the original data, the state transition probability was calculated, and the anomaly data were evaluated in quality. Taking the data of hydrologic sensor obtained from the Chu River as experimental data, experiments on the detection time and outlier detection performance were carried out respectively. The results show that the millions of data calculation by two slaves costs more time than that by one slave, but when calculating the tens of milllions of data, the time costed by two slaves is less than that by one slave, and the maximum reduction is 16.21%. The sensitivity of the evaluation is increased from 5.24% to 92.98%. It shows that under big data platform, the proposed algorithm which is based on the characteristics of hydrological data and combines forecast test and cluster test can effectively improve the computational efficiency of hydrologic time series detection for tens of millions data and has a significant improvement in sensitivity.

In recent years, the Generative Adversarial Network (GAN) has achieved remarkable success in text-to-image synthesis, but there are still problems such as edge blurring of images, unclear local textures, small sample variance. In view of the above shortcomings, based on Stack Generative Adversarial Network model (StackGAN++), a Multi-Level structure Generative Adversarial Networks (MLGAN) model was proposed, which is composed of multiple generators and discriminators in a hierarchical structure. Firstly, hierarchical structure coding method and word vector constraint were introduced to change the condition vector of generator of each level in the network, so that the edge details and local textures of the image were clearer and more vivid. Then, the generator and the discriminator were jointed by trained to approximate the real image distribution by using the generated image distribution of multiple levels, so that the variance of the generated sample became larger, and the diversity of the generated sample was increased. Finally, different scale images of the corresponding text were generated by generators of different levels. The experimental results show that the Inception scores of the MLGAN model reached 4.22 and 3.88 respectively on CUB and Oxford-102 datasets, which were respectively 4.45% and 3.74% higher than that of StackGAN++. The MLGAN model has improvement in solving edge blurring and unclear local textures of the generated image, and the image generated by the model is closer to the real image.