The fake speech produced by modern speech synthesis and timbre conversion systems poses a serious threat to the automatic speaker recognition system. Most of the existing fake speech detection systems perform well for the known attack types in the training process， but degrades significantly in detecting unknown attack types in practical applications. Therefore， combined with the recently proposed Dual?Path Res2Net （DP?Res2Net）， a semi?supervised end?to?end fake speech detection method based on time?domain waveforms was proposed. Firstly， semi?supervised learning was adopted for domain transfer to reduce the difference of data distribution between training set and test set. Then， for feature engineering， time-domain sampling points were input into DP?Res2Net directly， which increased the local multi?scale information and made full use of the dependence between audio segments. Finally， the embedded tensors were obtained to judge fake speech from natural speech after the input features going through the shallow convolution module， feature fusion module and global average pooling module. The performance of the proposed method was evaluated on the publicly available ASVspoof 2021 Speech Deep Fake evaluation set as well as the dataset VCC （Voice Conversion Challenge）. Experimental results show that the Equal Error Rate （EER） of the proposed method is 19.97%， which is 10.8% less than that of the official optimal baseline system， verifying that the semi?supervised end?to?end fake speech detection method based on time?domain waveforms is effective when recognizing unknown attacks and has higher generalization capability.

Image fusion technology plays an important role in computer-aided diagnosis. Detail extraction and energy preservation are two key issues in image fusion, and the traditional fusion methods address them simultaneously by designing the fusion method. However, it tends to cause information loss or insufficient energy preservation. In view of this, a fusion method was proposed to solve the problems of detail extraction and energy preservation separately. The first part of the method aimed at detail extraction. Firstly, the Non-Subsampled Shearlet Transform (NSST) was used to divide the source image into low-frequency and high-frequency subbands. Then, an improved energy-based fusion rule was used to fuse the low-frequency subbands, and an strategy based on the intuitionistic fuzzy set theory was proposed for the fusion of the high-frequency subbands. Finally, the inverse NSST was employed to reconstruct the image. In the second part, an intensity enhancement method was proposed for energy preservation. The proposed method was verified on 43 groups of images and compared with other eight fusion methods such as Principal Component Analysis (PCA) and Local Laplacian Filtering (LLF). The fusion results on two different categories of medical image fusion (Magnetic Resonance Imaging (MRI) and Positron Emission computed Tomography (PET), MRI and Single-Photon Emission Computed Tomography (SPECT)) show that the proposed method can obtain more competitive performance on both visual quality and objective evaluation indicators including Mutual Information (MI), Spatial Frequency (SF), Q value, Average Gradient (AG), Entropy of Information (EI), and Standard Deviation (SD), and can improve the quality of medical image fusion.

Concerning the low accuracy and long running time of link prediction between dynamic network nodes, a dynamic network representation learning model using denoising AutoEncoder (dAE) as the framework and combining with Graph Convolutional Network (GCN) and Long Short-Term Memory (LSTM) network, named dynGAELSTM, was proposed. Firstly, the GCN was used in the front-end of this model to capture the feature information of the high-order graph neighborhood of the dynamic network nodes. Secondly, the extracted information was input into the coding layer of the dAE to obtain the low-dimensional feature vectors, and the spatio-temporal dependent features of the dynamic network were obtained on the LSTM network. Finally, a loss function was constructed by comparing the prediction map reconstructed through the decoding layer of the dAE with the real map, so as to optimize the model to complete the link prediction. Theoretical analysis and simulation experiments showed that compared with the model with the second-best prediction performance, the dynGAELSTM model had the prediction performance improved by 0.79, 1.19 and 3.13 percentage points respectively, and the running time reduced by 0.92% and 1.73% respectively. In summary, the dynGAELSTM model has higher accuracy and lower complexity in the link prediction tasks compared to the existing models.

Aiming at the problem of express freight scheme decision under multiple uncertain factors, the express freight scheme decision model and decision rule based on intuitionistic fuzzy rough set over two universes were proposed. Based on the intuitionistic fuzzy rough set theory over two universes, a fuzzy approximate space over two universes for express freight scheme decision was determined. The consumption degrees of fixed cost, transportation cost, transfer cost, carbon emission, transfer time and other transportation indices were regarded as intuitionistic fuzzy numbers, and the intuitionistic fuzzy relation between evaluation indices and transportation schemes were used to calculate the lower approximation set and upper approximation set, and the maximum intuitionistic index and Hamming closeness degree were introduced to determine the transportation scheme decision rules. Taking an express freight transportation line from Lanzhou to Beijing as the example, the optimal transportation scheme was selected from the 9 modes of transportation combined by road, ordinary speed railway and air according to the decision rules. Sensitivity analysis of transportation cost and transfer cost was performed to verify the accuracy of the results. The two optimal transportation schemes finally selected show the applicability of the intuitionistic fuzzy rough set over two universes on such problems.

A stochastic optimization method was proposed in order to solve the vehicle number allocation problem of the minimum autonomous vehicle fleet driven by spatio-temporal multi-tasks of terminal delivery. Firstly, the influence of service time and waiting time on the route planning of autonomous vehicle fleet was analyzed to build the shortest route model, and the service sequence network was constructed based on the two-dimensional spatio-temporal network. Then, the vehicle number allocation problem of the minimum autonomous vehicle fleet was converted into a network maximum flow problem through the network transformation, and a minimum fleet model was established with the goal of minimizing the vehicle number of the fleet. Finally, the Dijkstra-Dinic algorithm combining Dijkstra algorithm and Dinic algorithm was designed according to the model features in order to solve the vehicle number allocation problem of the minimum autonomous vehicle fleet. Simulation experiments were carried out in four different scales of service networks, the results show that:under different successful service rates, the minimum size of autonomous vehicle fleet is positively correlated with the scale of service network, and it decreases with the increase of waiting time and gradually tends to be stable, the One-stop operator introduced into the proposed algorithm greatly improves the search efficiency, and the proposed model and algorithm are suitable for the calculation of the minimum vehicle fleet in large-scale service network.

In order to solve the training difficulty of small sample data in deep learning and increase the training efficiency of DCGAN (Deep Convolutional Generative Adversarial Network), an improved DCGAN algorithm was proposed to perform the augmentation of small sample data. In the method, Wasserstein distance was used to replace the loss model in the original model at first. Then, spectral normalization was added in the generation network, and discrimination network to acquire a stable network structure. Finally, the optimal noise input dimension of sample was obtained by the maximum likelihood estimation and experimental estimation, so that the generated samples became more diversified. Experimental result on three datasets MNIST, CelebA and Cartoon indicated that the improved DCGAN could generate samples with higher definition and recognition rate compared to that before improvement. In particular, the average recognition rate on these datasets were improved by 8.1%, 16.4% and 16.7% respectively, and several definition evaluation indices on the datasets were increased with different degrees, suggesting that the method can realize the small sample data augmentation effectively.

Focusing on the high execution cost of star-join between the fact table and multiple dimension tables in On-line Analytical Processing (OLAP), a star-join optimization technique was proposed for advanced multicore CPU (Central Processing Unit) and GPU (Graphics Processing Unit). Firstly, the vector index based vectorized star-join algorithm on CPU and GPU platforms was proposed for the intermediate materialization cost problem in star-join in multicore CPU and GPU platforms. Secondly, the star-join operation based on vector granularity was presented according to the vector division for CPU cache size and GPU shared memory size, so as to optimize the vector index materialization cost in star-join. Finally, the compressed vector index based star-join algorithm was proposed to compress the fixed-length vector index to the variable-length binary vector index, so as to improve the storage access efficiency of the vector index in cache under low selection rate. Experimental results show that the vectorized star-join algorithm achieves more than 40% performance improvement compared to the traditional row-wise or column-wise star-join algorithms on multicore CPU platform, and the vectorized star-join algorithm achieves more than 15% performance improvement compared to the conventional star-join algorithms on GPU platform; in the comparison with the mainstream main-memory databases and GPU databases, the optimized star-join algorithm achieves 130% performance improvement compared to the optimal main-memory database Hyper, and achieves 80% performance improvement compared to the optimal GPU database OmniSci. It can be seen that the vector index based star-join optimization technique effectively improves the multiple table join performance, and compared with the traditional optimization techniques, the vector index based vectorized processing improves the data storage access efficiency in small cache, and the compressed vector further improves the vector index access efficiency in cache.

Concerning the problems that the traditional Cartesian Genetic Programming (CGP) is lack of diversity of mutation operation and the evolutionary strategy used in it has limitations, an ADvanced Real-Valued Cartesian Genetic Programming algorithm based on quasi-oppositional mutation (AD-RVCGP) was proposed. Firstly, the 1+lambda evolutionary strategy was adopted in the evolution process in AD-RVCGP just like in the traditional CGP, that is lambda offsprings were generated by a parent only through mutation operation. Secondly, three mutation operators including quasi-oppositional mutation operator, terminal mutation operator and single-point mutation operator were dynamically selected in the process of evolution, and the information of oppositional individuals was used for the mutation operation. Finally, in the evolution process, different parents were selected in the algorithm to generate the next generation individuals according to the state of evolution stage. In the test of symbolic regression problem, the convergence speed of the proposed AD-RVCGP was about 30% faster than that of the traditional CGP, and the running time was about 20% less. In addition, the error between the optimal solution obtained by AD-RVCGP and the real optimal solution was smaller than the optimal solution obtained by the traditional CGP and the real optimal solution. Experimental results show that the proposed AD-RVCGP has high convergence speed and precision for solving problem.

The existing Nakagami parametric imaging of ultrasonic harmonic envelope signals can realize non-invasive monitoring of the ablation process, but it cannot estimate the ablation area accurately. In order to solve the problem, a Gaussian Approximation adaptive Threshold Segmentation (GATS) method based on ultrasonic harmonic envelope Nakagami parameter images was proposed to monitor microwave ablation areas accurately and effectively. Firstly, a high-pass filter was used to obtain the harmonic components of the ultrasound echo Radio Frequency (RF) signal. Then, the Nakagami shape parameters of the harmonic signal envelope were estimated, and Nakagami parameter image was generated by composite window imaging. Finally, Gaussian approximation of Nakagami parameter image was applied to present the ablation area, the anisotropic smoothing preprocessing was performed to the approximated image, and the threshold segmentation of the smoothed image was used to accurately estimate the ablation area. The results of microwave ablation experiments show that, the long and short axis errors of the threshold segmentation ablation area after anisotropic smoothing based on Perona-Malik (P-M) algorithm and the actual ablation area are reduced by 3.15 percentage points and 2.21 percentage points respectively compared with the errors obtained by using Catte algorithm, and decreased by 7.87 percentage points and 5.74 percentage points compared with the errors obtained by using Median algorithm. It can be seen that GATS using P-M algorithm for ultrasonic harmonic envelope Nakagami parameter images can estimate the ablation area more accurately and provide effective monitoring for clinical ablation surgery.

Under the development of modern pharmaceutical and computer technologies, using artificial intelligence technology to accelerate drug development progress has become a research hotspot. And efficient prediction of organic reaction products is a key issue in drug retrosynthesis path planning. Concerning the problem of uneven distribution of chemical reaction types in the sample dataset, an Active Sampling-training Gated Graph Convolutional Neural-network (ASGGCN) model was proposed. Firstly, the SMILES (Simplified Molecular Input Line Entry Specification) codes of the chemical reactants were input into the model, and the location of the reaction center was predicted through Gated Graph Convolutional Neural-network (GGCN) and attention mechanism. Then, according to chemical constraint conditions and the candidate reaction centers, the possible chemical bond combinations were enumerated to generate candidate reaction products. After that, the gated graph convolutional difference network was used to rank the candidate products and obtain the final reaction product. Compared with the traditional graph convolutional network, the gated graph convolutional network has three weight parameter matrices and fuse the information through gating, so it can obtain more abundant atom hidden feature information. At the same time, the gated graph convolutional network is trained by active sampling, which can take into account both the analysis abilities of poor samples and ordinary samples. Experimental results show that the Top-1 prediction accuracy of the reaction product of the proposed model reaches 87.2%, which is increased by 1.6 percentage points compared to the accuracy of WLDN (Weisfeiler-Lehman Difference Network) model, illustrating that the organic reaction products can be predicted more accurately by the proposed model.

With substantial attention being received from industry and academia over last decade, salient object detection has become an important fundamental research in computer vision. The solution of salient object detection will be helpful to make breakthroughs in various visual tasks. Although various works have achieved remarkable success for saliency detection tasks in visible light scenes, there still remain a challenging issue on how to extract salient objects with clear boundary and accurate internal structure in weak light images with low signal-to-noise ratios and limited effective information. For that fuzzy boundary and incomplete internal structure cause low accuracy of salient object detection in weak light scenes, an Ant Colony Optimization (ACO) algorithm based saliency detection framework was proposed. Firstly, the input image was transformed into an undirected graph with different nodes by multi-scale superpixel segmentation. Secondly, the optimal feature selection strategy was adopted to capture the useful information contained in the salient object and eliminate the redundant noise information from weak light image with low contrast. Then, the spatial contrast strategy was introduced to explore the global saliency cues with relatively high contrast in the weak light image. To acquire more accurate saliency estimation at low signal-to-noise ratio, the ACO algorithm was used to optimize the saliency map. Through the experiments on three public datasets (MSRA, CSSD and PASCAL-S) and the Nighttime Image (NI) dataset, it can be seen that the Area Under the Curve (AUC) value of the proposed model reached 87.47%, 84.27% and 81.58% on three public datasets respectively, and the AUC value of the model was increased by 2.17 percentage points compared to that of the Low Rank Matrix Recovery (LR) model (which ranked the second) on the NI dataset. The results demonstrate that the proposed model has the detection effect with more accurate structure and clearer boundary compared to 11 mainstream saliency detection models and effectively suppresses the interference of weak light scenes on the detection performance of salient objects.

The differences between different protocols are not considered when selecting identification features in the existing statistical feature-based identification methods. In order to solve the problem, a Semi-supervised Subspace-clustering Protocol Identification Approach (SSPIA) was proposed by combining semi-supervised learning and Fuzzy Subspace Clustering (FSC) method. Firstly, the prior constraint condition was obtained by transforming the labeled sample flow into pairwise constraints information. Secondly, the Semi-supervised Fuzzy Subspace Clustering (SFSC) algorithm was proposed on this basis and was used to guide the process of subspace clustering by using the constraint condition. Then, the mapping between class clusters and protocol types was established to obtain the weight coefficient of each protocol feature, and an individualized cryptographic protocol feature library was constructed for subsequent protocol identification. Finally, the clustering effect and identification effect experiments of five typical cryptographic protocols were carried out. Experimental results show that, compared with the traditional K-means method and FSC method, the proposed SSPIA has better clustering effect, and the protocol identification classifier constructed by SSPIA is more accurate, has higher protocol identification rate and lower error identification rate. The proposed SSPIA improves the identification effect based on statistical features.

With the continuous popularization and promotion of social networks, compared with traditional text description, people are inclined to post reviews with both images and texts to express their feelings and opinions. The existing visual-textual sentiment analysis methods only consider the high-level semantic relation between images and texts, but pay less attention to the correlation between the low-level visual features and middle-level aesthetic features of images and the sentiment of texts. Thus, a visual-textual sentiment analysis method based on Multi-Level Spatial Attention (MLSA) was proposed. In the proposed method, driven by text content, MLSA was used to design the feature fusion method between images and texts. This feature fusion method not only focused on the image entity features related to texts, but also made full use of the middle-level aesthetic features and low-level visual features of images, so as to to mine the sentiment co-occurrence between images and texts from various perspectives. Compared to the classification effect of the best method among the comparison methods, the classification effect of the model was improved by 0.96 and 1.06 percentage points on accuracy, and improved by 0.96 and 0.62 percentage points on F1 score on two public multimodal sentiment datasets (MVSA_Single and MVSA_Multi) respectively. Experimental results show that the comprehensive analysis of the hierarchical relationship between text features and image features can effectively enhance the neural network's ability to capture the emotional semantics of texts and images, so as to predict the overall sentiment of texts and images more accurately.

In order to solve the problem of low accuracy of video-based person re-identification caused by factors such as occlusion, background interference, and person appearance and posture similarity in video surveillance, a video-based person re-identification method of Evenly Sampling-random Erasing (ESE) and global temporal feature pooling was proposed. Firstly, aiming at the situation where the object person is disturbed or partially occluded, a data enhancement method of evenly sampling-random erasing was adopted to effectively alleviate the occlusion problem, improving the generalization ability of the model, so as to more accurately match the person. Secondly, to further improve the accuracy of video-based person re-identification, and learn more discriminative feature representations, a 3D Convolutional Neural Network (3DCNN) was used to extract temporal and spatial features. And a Global Temporal Feature Pooling (GTFP) layer was added to the network before the output of person feature representations, so as to ensure the obtaining of spatial information of the context, and refine the intra-frame temporal information. Lots of experiments conducted on three public video datasets, MARS, DukeMTMC-VideoReID and PRID-201l, prove that the method of jointing evenly sampling-random erasing and global temporal feature pooling is competitive compared with some state-of-the-art video-based person re-identification methods.

Concerning the low tag recognition throughput caused by frame length limitation in the Dynamic Frame Slot Aloha (DFSA) algorithm, a Logistic mapping based DFSA (Logistic-DFSA) algorithm was proposed. First, the sequence generated by logistic mapping was used as the spreading code, and the spread spectrum technology was combined with the DFSA algorithm to realize the parallel recognition of multiple tags with one slot. Second, the influence of frame length, spreading code length and the number of tags on throughput in the recognition process was analyzed, and the optimal frame length and spreading code length were obtained. Finally, based on the number of remaining tags after a frame, a repeating frame algorithm with all tags recognizable was proposed. Simulation results show that compared with the DFSA algorithm, the Logistic-DFSA algorithm has reduced the total number of slots for tag recognition by 98.3% and increased the system throughout by 162%. Therefore, the Logistic-DFSA algorithm can greatly reduce the total number of slots, improve the system throughput, and effectively identify tags within the range of the reader.

Aiming at the difficulty of finding terrorist attack organizations and the imbalance of terrorist attack data samples, a terrorist attack organization prediction method based on feature selection and hyperparameter optimization was proposed. First, by taking the advantage of Random Forest (RF) in dealing with imbalanced data, the backward feature selection was carried out through the RF iteration. Second, four mainstream classifiers including Decision Tree (DT), RF, Bagging and XGBoost were used to classify and predict terrorist attack organizations, and the Bayesian optimization method was used to optimize the hyperparameters of these classifiers. Finally, the Global Terrorism Database (GTD) was used to evaluate the classification prediction performance of these classifiers on the majority class samples and minority class samples. Experimental results show that the proposed method improves the classification and prediction performance of terrorist attack organizations, and the classification and prediction performance is the best when using RF and Bagging, with the accuracy of 0.823 9 and 0.831 6 respectively. Especially for minority class samples, the classification and prediction performance when using RF and Bagging is significantly improved.

The low-rank constraint model based on traditional Nuclear Norm Minimization(NNM)tends to cause local texture detail loss in the denoising of Low-Dose CT(LDCT)image. To tackle this issue,a regional-content-aware weighted NNM algorithm was proposed for LDCT image denoising. Firstly,a Singular Value Decomposition(SVD)based method was proposed to estimate the local noise intensity in LDCT image. Then,the target image block matching was performed based on the local statistical characteristics. Finally,the weights of the nuclear norms were adaptively set based on both the local noise intensity of the image and the different singular value levels,and the weighted NNM based LDCT image denoising was realized. The simulation results illustrated that the proposed algorithm decreased the Root Mean Square Error(RMSE)index by 30. 11%,14. 38% and 8. 75% respectively compared with the traditional NNM,total variation minimization and transform learning algorithms,and improved the Structural SIMilarity(SSIM)index by 34. 24%,23. 06% and 11. 52% respectively compared with the above three algorithms. The experimental results on real clinical data illustrated that the mean value of the radiologists' scores of the results obtained by the proposed algorithm was 8. 94,which is only 0. 21 lower than that of the corresponding full dose CT images,and was significantly higher than those of the traditional NNM,total variation minimization and transform learning algorithms. The simulation and clinical experimental results indicate that the proposed algorithm can effectively reduce the artifact noise while preserving the texture detail information in LDCT images.

In order to solve the intelligence extraction of information in the era of remote sensing big data,it is important to build the model and method of intelligent information analysis fitting the intrinsic characteristics of remote sensing data. To meet the demand of universal remote sensing intelligent acquisition of large-scale water information,an intelligent extraction method of remote sensing water information based on visual selective attention mechanism and AdaBoost algorithm was proposed. Firstly,by the optimization design of RGB color scheme of remote sensing multi-feature index,the enhancement and visual representation of the water information image features were realized. Then,in HSV color space,the key node information of the chromatic aberration distance image was used to construct the classification feature set,and AdaBoost algorithm was used to construct the water recognition classifier. On this basis,the category that the water belongs to was automatically recognized from the image color clustering result,so as to realize the intelligent extraction of water information. Experimental results show that the proposed method has the water information extraction results improved on Leak Rate(LR)and Composite Classification Accuracy(CCA). At the same time,the proposed method not only effectively reduces the dependence on high quality training samples,but also has good performance on the recognition of temporary water areas such as water with high sediment concentration at wet season and submerged area caused by flooding.

It is difficult to obtain the overall optimal solution by the traditional order batching and the picking path step-by-step optimization of picking operation in the logistics distribution center. In order to improve the efficiency of picking operation, a joint picking strategy based on nested genetic algorithm for order batching and path optimization was proposed. Firstly, the joint optimization model of order batching and picking path was established with the shortest total picking time as the objective function. Then, a nested genetic algorithm was designed to solve the model with the consideration of the complexity of double optimizations. The order batching result was continuously optimized in the outer layer, and the picking path was optimized in the inner layer according to the order batching result in the outer layer. Results of the examples show that, compared with the traditional strategies of order step-by-step optimization and step-by-step optimization in batches, the proposed strategy has reduced the picking time by 45.6% and 6% respectively, and the joint optimization model based on nested genetic algorithm results in shorter picking path and less picking time. To verify that the proposed algorithm has better performance on orders with different sizes, the simulation experiments were performed to the examples with 10, 20, 50 orders respectively. The results show that, with the increase of order quantity, the overall picking distance and time are further reduced, the decrease of picking time is risen from 6% to 7.2%.The joint optimization model of picking operation based on nested genetic algorithm and its solution algorithm can effectively solve the joint optimization problem of order batching and picking path, and provide the basis for the optimization of picking system in the distribution center.

Focusing on the problem of chess piece positioning error in the chess robot system under uneven illumination distribution, a general chess piece positioning method based on block convex hull detection and image mask was proposed. Firstly, the set of points on the outline of the chessboard were extracted, the coordinates of the four vertices of the chessboard were detected using the block convex hull method. Secondly, the coordinates of the four vertices of the chessboard in the standard chessboard image were defined, and the transformation matrix was calculated by the perspective transformation principle. Thirdly, the type of the chessboard was recognized based on the difference between the small square areas of different chessboards. Finally, the captured chessboard images were successively corrected to the standard chessboard images, and the difference images of two adjacent standard chessboard images were obtained, then the dilation, image mask multiplication and erosion operations were performed on the difference images in order to obtain the effective areas of chess pieces and calculate their center coordinates. Experimental results demonstrate that, the proposed method has the average positioning accuracy of Go and Chinese chess pieces arrived by 95.5% and 99.06% respectively under four kinds of uneven illumination conditions, which are significantly improved in comparison with other chess piece positioning algorithms. At the same time, the proposed method can solve the inaccurate local positioning problem of chess pieces caused by adhesion of chess pieces, chess piece projection and lens distortion.

The manual segmentation of chromosome images is time-consuming and laborious, and the accuracy of current automatic segmentation methods is not poor. Therefore, based on improved Mask R-CNN (Mask Region-based Convolutional Neural Network), a chromosome image segmentation framework named Mask Oriented R-CNN (Mask Oriented Region-based Convolutional Neural Network) was proposed, which introduced orientation information to perform instance segmentation of chromosome images. Firstly, the regression branch of oriented bounding boxes was added to predict the compact bounding boxes and obtain orientation information. Secondly, a novel Intersection-over-Union (IoU) metric called AwIoU (Angle-weighted Intersection-over-Union) was proposed to improve the criterion of redundant bounding boxes by combining the relationship between the orientation information and edges. Finally, the oriented convolutional path structure was realized to reduce the interference in mask prediction by copying the path of mask branch and selecting the training path according to the orientation information of the instances. Experimental results show that compared with the baseline model Mask R-CNN, Mask Oriented R-CNN has the mean average precision increased by 10.22 percentage points when the IoU threshold is 0.5, and the mean metric increased by 4.91 percentage points when the IoU threshold is from 0.5 to 0.95. Experimental results show that the Mask Oriented R-CNN framework achieves better segmentation results than the baseline model in chromosome image segmentation, which is helpful to achieve automatic segmentation of chromosome images.

In the edge computing, the communication between edge computing nodes and terminal devices is vulnerable to impersonation attacks, therefore a physical-layer impersonation attack detection algorithm based on Deep Q-Network (DQN) was proposed. Firstly, an impersonation attack model was built in the edge computing network, a hypothesis test based on the physical-layer Channel State Information (CSI) was established by the receiver, and the Euclidean distance between the currently measured CSI and the last recorded CSI was taken as the test statistics. Secondly, for the dynamic environment of edge computing, the DQN algorithm was used to adaptively select the optimal test threshold with the goal of maximizing the gain of the receiver. Finally, whether the current sender was an impersonation attacker was determined by comparing the statistics with the test threshold. The simulation results show that the Signal-to-Interference plus Noise Ratio (SINR) and channel gain ratio have certain effect on the performance of the detection algorithm, but when the relative change of channel gain is lower than 0.2, the false alarm rate, miss rate and average error rate of the algorithm are less than 5%. Therefore, the detection algorithm is adaptive to the dynamical environment of edge computing.

The current research on the activity semantic recognition only extracts the sequence features and periodic features on the time dimension, and lacks deep mining of spatial information. To solve these problems, an activity semantic recognition method based on joint features and eXtreme Gradient Boosting (XGBoost) was proposed. Firstly, the activity periodic features in the temporal information as well as the latitude and longitude features in the spatial information were extracted. Then the latitude and longitude information was used to extract the heat features of the spatial region based on the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm. The user activity semantics was represented by the feature vectors combined with these features. Finally, the activity semantic recognition model was established through the XGBoost algorithm in the integrated learning method. On two public check-in datasets of FourSquare, the model based on joint features has a 28 percentage points improvement in recognition accuracy compared to the model with only temporal features, and compared with the Context-Aware Hybrid (CAH) method and the Spatial Temporal Activity Preference (STAP) method, the proposed method has the recognition accuracy increased by 30 percentage points and 5 percentage points respectively. Experimental results show that the proposed method is more accurate and effective on the problem of activity semantic recognition compared to the the comparison methods.

Aiming at the problem of low reconstruction quality of the existing Compressed Sensing Magnetic Resonance Imaging (CSMRI) algorithms at low sampling rates, an imaging method combining deep priors and non-local similarity was proposed. Firstly, a deep denoiser and Block Matching and 3D filtering (BM3D) denoiser were used to construct a sparse representation model that can fuse multiple priori knowledge of images. Secondly, the undersampled k-space data was used to construct a compressed sensing magnetic resonance imaging optimization model. Finally, an alternative optimization method was used to solve the constructed optimization problem. The proposed algorithm can not only use the deep priors through the deep denoiser, but also use the non-local similarity of the image through the BM3D denoiser to reconstruct the image. Compared with the reconstruction algorithms based on BM3D, experimental results show that the proposed algorithm has the average peak signal-to-noise ratio of reconstruction increased about 1 dB at the sampling rates of 0.02, 0.06, 0.09 and 0.13. Compared with the existing MRI algorithm WaTMRI (Magnetic Resonance Imaging with Wavelet Tree sparsity),DLMRI (Dictionary Learning for Magnetic Resonance Imaging), DUMRI-BM3D (Magnetic Resonance Imaging based on Dictionary Updating and Block Matching and 3D filtering), etc, the images reconstructed by the proposed algorithm contain a lot of texture information, which are the closest to the original images.

Most of the existing automatic emotion annotation methods only extract the single recognition feature from acoustic layer or language layer. While Yi language is affected by the factors such as too many branch dialects and high complexity, so the accuracy of automatic annotation of Yi language with single-layer emotion feature is low. Based on the features such as rich emotional affixes in Yi language, a double-layer feature fusion method was proposed. In the method, the emotional features from acoustic layer and language layer were extracted respectively, the methods of generating sequence and adding units as needed were applied to complete the feature sequence alignment, and the process of automatic emotion annotation was realized through the corresponding feature fusion and automatic annotation algorithm. Taking the speech and text data of Yi language in a poverty alleviation log database as samples, three different classifiers were used for comparative experiments. The results show that the classifier has no obvious effect on the automatic annotation results, and the accuracy of automatic annotation after the fusion of double-layer features is significantly improved, the accuracy is increased from 48.1% of acoustic layer and 34.4% of language layer to 64.2% of double-layer fusion.

Focused on the issue that the traditional interest area based visualization method can not pay attention to the details in the process of analyzing pilot eye movement data, a visual analysis method of eye movement data based on user-defined interest area was proposed. Firstly, according to the specific analysis task, the self-divison and self-definition of the background image of the task were introduced. Then, multiple auxiliary views and interactive approaches were combined, and an eye movement data visual analysis system for pilot training was designed and implemented to help analysts analyze the difference of eye movement between different pilots. Finally, through case analysis, the effectiveness of the visual analysis method and the practicability of the analysis system were proved. The experimental results show that compared with the traditional method, in the proposed method, the analysts' initiative in the analysis process is increased. The analysts are allowed to explore the local details of the task background in both global and local aspects, making the analysts' analyze the data in multi-angle; the analysts are allowed find the flight students' cognitive difficulties in the training process as a whole, so as to develop more targeted and more effective training courses.

In order to reduce the delay time of disaster relief materials distribution and the total transportation time of disaster relief vehicles, the concept of urgency was introduced to establish a vehicle routing problem model of disaster relief vehicles based on urgency, and an improved Genetic Algorithm (GA) was designed to solve the model. Firstly, multiple strategies were used to generate the initial population. Then, an urgency-based task redistribution algorithm was proposed as local search operator. The proposed algorithm achieved the optimal delay time and total transportation time based on urgency. The delay time was reduced by rescheduling the vehicle or adjusting the delivery sequence for delay placements. The routes of the vehicles without delay were optimized to reduce the total transportation time. In the experiments, the proposed algorithm was compared with First-Come-First-Served (FCFS) algorithm, Sort by URGency (URGS) and GA on 17 datasets. Results show that the Genetic Algorithm with Task Redistribution strategy based on Urgency Degree (TRUD-GA) reduces the average delay time by 25.0% and decreases the average transportation time by 1.9% compared with GA, and has more obvious improvement compared with FCFS and URGS algorithms.

An image similarity algorithm based on Lance Williams distance was proposed to solve the problem that the boundary of similarity between dust image and dust-free image is not obvious when illumination changes in desktop dust detection. The Lance Williams distance between template image and the images with or without dust was converted to the similarity value of (0, 1] and the difference of similarity values was expanded with exponential function properties in the algorithm. In order to enhance the dust texture feature information, the gray image was convolved with the Laplacian and then the feature parameters were obtained using co-occurrence matrix feature extraction algorithm and combined into a one-dimensional vector. The similarity of feature parameter vectors between template image and to-be-detected image was calculated by the improved similarity algorithm to determine whether the desktop has dust or not. Experimental results show that the similarity is more than 90.01% between dust-free images and less than 62.57% between dust and dust-free images in the range of 300~900 lux illumination. The average of the two similarities can be regarded as the threshold to determine whether the desktop has dust or not when illumination changes.

Multi-exposure image fusion technology directly combines a sequence of images with the same scene but different exposure levels into a high-quality image with more details of scene. Aiming at the problems of poor local contrast difference and color distortion of existing algorithms, a new multi-exposure image fusion algorithm was proposed based on Retinex theoretical model. Firstly, based on Retinex theoretical model, the exposure sequence images were divided into an illumination component sequence and a reflection component sequence by using the illumination estimation algorithm, and then two sets of sequences were processed by different fusion methods. For the illumination component, the variation characteristics of global brightness of scene were guaranteed and the effects of overexposed and underexposed regions were weakened, while for the reflection component, the evaluation parameters of moderate exposure were used to better preserve the color and detail information of scene. The proposed algorithm was analyzed from both subjective and objective aspects. The experimental results show that compared with traditional algorithm based on image domain synthesis, the proposed algorithm has an average increase of 1.7% in Structural SIMilarity (SSIM) and has better effect in the processing of image color and local details.

Focusing on the issues of low quality of image caption, insufficient utilization of image features and single-level structure of recurrent neural network in image caption generation, an image caption generation algorithm based on multi-attention and multi-scale feature fusion was proposed. The pre-trained target detection network was used to extract the features of the image from the convolutional neural network, which were input into the multi-attention structures at different layers. Each attention part with features of different levels was related to the multi-level recurrent neural networks sequentially, constructing a multi-level image caption generation network model. By introducing residual connections in the recurrent networks, the network complexity was reduced and the network degradation caused by deepening network was avoided. In MSCOCO datasets, the BLEU-1 and CIDEr scores of the proposed algorithm can achieve 0.804 and 1.167, which is obviously superior to top-down image caption generation algorithm based on single attention structure. Both artificial observation and comparison results velidate that the image caption generated by the proposed algorithm can show better details.