Aiming at the problems of low accuracy of driving command prediction, bulky model structure and a large amount of information redundancy in existing end-to-end autonomous driving methods, a new end-to-end autonomous driving model based on deep visual attention neural network was proposed. In order to effectively extract features of autonomous driving scenes, a deep visual attention neural network, which is composed of the convolutional neural network, the visual attention layer and the long short-term memory network, was proposed by introducing a visual attention mechanism into the end-to-end autonomous driving model. The proposed model was able to effectively extract spatial and temporal features of driving scene images, focus on important information and reduce information redundancy for realizing the end-to-end autonomous driving that predicts driving commands from sequential images input by front-facing camera. The data from a simulated driving environment were used for training and testing. The root mean square errors of the proposed model for prediction of the steering angle in four scenes including country road, highway, tunnel and mountain road are 0.009 14, 0.009 48, 0.002 89 and 0.010 78 respectively, which are all lower than the results of the method proposed by NVIDIA and the method based on the deep cascaded neural network. Moreover, the proposed model has fewer network layers compared with the networks without the visual attention mechanism.

To solve the problems of requiring a large number of samples, not associating with time information, and not using global navigation information in motion planning for autonomous driving based on machine learning, a motion planning method for autonomous driving with directional navigation based on deep spatio-temporal Q-network was proposed. Firstly, in order to extract the spatial features in images and the temporal information between continuous frames for autonomous driving, a new deep spatio-temporal Q-network was proposed based on the original deep Q-network and combined with the long short-term memory network. Then, to make full use of the global navigation information of autonomous driving, directional navigation was realized by adding the guide signal into the images for extracting environment information. Finally, based on the proposed deep spatio-temporal Q-network, a learning strategy oriented to autonomous driving motion planning model was designed to achieve the end-to-end motion planning, where the data of steering wheel angle, accelerator and brake were predicted from the input sequential images. The experimental results of training and testing results in the driving simulator named Carla show that in the four test roads, the average deviation of this algorithm is less than 0.7 m, and the stability performance of this algorithm is better than that of four comparison algorithms. It is proved that the proposed method has better learning performance, stability performance and real-time performance to realize the motion planning for autonomous driving with global navigation route.

To handle the issues of lacking paired training samples and inconsistency between frames in translation from virtual to real driving scenes, a video translation model based on Generative Adversarial Networks was proposed in this paper. In order to solve the problem of lacking data samples, the model adopted a “dual networks” architecture, where the semantic segmentation scene was used as an intermediate transition to build front-part and back-part networks, respectively. In the front-part network, a convolution network and a deconvolution network were adopted, and the optical flow network was also used to extract the dynamic information between frames to implement continuous video translation from virtual to semantic segmentation scenes. In the back-part network, a conditional generative adversarial network was used in which a generator, an image discriminator and a video discriminator were designed and combined with the optical flow network to implement continuous video translation from semantic segmentation to real scenes. Data collected from an autonomous driving simulator and a public data set were used for training and testing. Virtual to real scene translation can be achieved in a variety of driving scenarios, and the translation effect is significantly better than the comparative algorithms. Experimental results show that the proposed model can handle the problems of the discontinuity between frames and the ambiguity for moving obstacles to obtain more continuous videos when applying in various driving scenarios.

In order to improve the detection level of bridge cracks,and solve the time-consuming and laborious problem in manual detection and the parameters to be set manually in traditional image processing methods,an improved bridge crack detection algorithm was proposed based on GoogLeNet. Firstly,a large-scale bridge crack Retinex-Laplace-Histogram equalization(RLH)dataset was constructed for model training and testing. Secondly,based on the original GoogLeNet model,the inception module was improved by using the normalized convolution kernel,three improved schemes were used to modify the beginning of the network,the seventh and later inception layers were removed,and a bridge crack feature image classification system was established. Finally,the sliding window was used to accurately locate the cracks and the lengths and widths of the cracks were calculated by the skeleton extraction algorithm. The experimental results show that compared with the original GoogLeNet network,the improve-GoogLeNet network increased the recognition accuracy by 3. 13%, and decreased the training time to the 64. 6% of the original one. In addition,the skeleton extraction algorithm can consider the trend of the crack,calculate the width more accurately,and the maximum width and the average width can be calculated. In summary,the classification and measurement method proposed in this paper have the characteristics of high accuracy,fast speed,accurate positioning and accurate measurement.

The scalar neurons of convolutional neural networks cannot express the feature location information,and have poor adaptability to the complex vehicle driving environment,resulting in low traffic sign recognition rate. Therefore,an intelligent traffic sign recognition method based on capsule network was proposed. Firstly,the very deep convolutional neural network was used to improve the feature extraction part. Then,a pooling layer was introduced in the main capsule layer. Finally,the movement index average method was used for improving the dynamic routing algorithm. The test results on the GTSRB dataset show that the improved capsule network method improves the recognition accuracy in special scenes by 10. 02 percentage points. Compared with the traditional convolutional neural network,the proposed method has the recognition time for single image decreased by 2. 09 ms. Experimental results show that the improved capsule network method can meet the requirement of accurate and real-time traffic sign recognition.

To improve the performance of vehicle classification by making full use of image information from traffic surveillance, Histogram of Oriented Gradient Convolutional (HOG-C) features extraction method was added on the capsule network, and a Capsule Network model fusing with HOG-C features (HOG-C CapsNet) was proposed. Firstly, the gradient data in the images were calculated by the gradient statistical feature extraction layer, and then the Histogram of Oriented Gradient (HOG) feature map was plotted. Secondly, the color information of the image was extracted by the convolutional layer, and then the HOG-C feature map was plotted with the extracted color information and HOG feature map. Finally, the HOG feature map was input into to the convolutional layer extract its abstract features, and the abstract features were encapsulated through a capsule network into capsules with the three-dimensional spatial feature representation, so as to realize the vehicle classification by dynamic routing algorithm. Compared with other related models on the BIT-Vehicle dataset, the proposed model has the accuracy of 98.17%, the Mean Average Precision (MAP) of 97.98%, the Mean Average Recall (MAR) of 98.42% and the comprehensive evaluation index of 98.20%. Experimental results show that the vehicle classification in traffic surveillance scenarios can be achieved with better performance by using HOG-C CapsNet.

Aiming at the problem that the deep learning network model applied to medical image target detection only has a fixed receptive field and cannot effectively detect the cervical lymph nodes with obvious morphological and scale differences, a new recognition algorithm based on adaptive receptive field mechanism was proposed, applying deep learning to the automatic recognition of cervical lymph nodes in complete three-dimensional medical images at the first time. Firstly, the semi-random sampling method was used to crop the medical sequence images to generate the grid-based local image blocks and the corresponding truth labels. Then, the DeepNode network based on the adaptive receptive field mechanism was constructed and trained through the local image blocks and labels. Finally, the trained DeepNode network model was used for prediction. By inputting the whole sequence images, the cervical lymph node recognition results corresponding to the whole sequence was obtained end-to-end and quickly. On the cervical lymph node dataset, the cervical lymph node recognition using the DeepNode network has the recall rate of 98.13%, the precision of 97.38%, and the number of false positives per scan is only 29, and the time consumption is relatively shorter. The analysis of the experimental results shows that compared with current algorithms such as the combination of two-dimensional and three-dimensional convolutional neural networks, the general three-dimensional object detection and the weak supervised location based recognition, the proposed algorithm can realize the automatic recognition of cervical lymph nodes and obtain the best recognition results. The algorithm is end-to-end, simple and efficient, easy to be extended to three-dimensional target detection tasks for other medical images and can be applied to clinical diagnosis and treatment.

To address the problems that rule-based motion planning algorithms under constraints need pre-definition of rules and temporal features are not considered in deep learning-based methods, a motion planning model based on deep cascading neural networks was proposed. In this model, the two classical deep learning models, Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network, were combined to build a novel cascaded neural network, the spatial and temporal features of the input images were extracted respectively, and the nonlinear relationship between the input sequential images and the output motion parameters were fit to achieve the end-to-end planning from the input sequential images to the output motion parameters. In experiments, the data of simulated environment were used for training and testing. Results show that the Root Mean Squared Error (RMSE) of the proposed model in four scenes including country road, freeway, tunnel and mountain road is less than 0.017, and the stability of the prediction results of the proposed model is better than that of the algorithm without using cascading neural network by an order of magnitude. Experimental results show that the proposed model can effectively learn human driving behaviors, eliminate the effect of cumulative errors and adapt to different scenes of a variety of road conditions with good robustness.

Since a large number of Trojans are easily tracable on the Internet, a new Trojan attack scheme based on multimedia document was proposed. Firstly, the Trojan program was embedded into a carrier image as secret data by steganography. After the Trojan program was successfully injected, the encrypted user information was also hidden into the carrier image by steganography. Then the host automatically uploaded pictures to a social network. Finally, the attacker downloaded images from the social network and extracted secret data from images. The theoretical analysis and simulation results show that the proposed JPEG image steganography algorithm has good performance, and the Trojan scheme based on it outperfoms some existing algorithms in concealment, anti-forensics, anti-tracking and penetrating auditing. Such Trojans in social networks can cause user privacy leaks, so some precautions are given at last.

Focusing on the problem of poor memory in long sentences and the lack of core words' influence in semantic relation extraction, an Attention based bidirectional Neural Turing Machine (Ab-NTM) model was proposed. Instead of a Recurrent Neural Network (RNN), a Neural Turing Machine (NTM) was used firstly, and a Long Short-Term Memory (LSTM) network was acted as a controller, which contained larger and non-interfering storage, and it could hold longer memories than the RNN. Secondly, an attention layer was used to organize the context information on the word level so that the model could pay attention to the core words in sentences. Finally, the labels were gotten through the classifier. Experiments on the SemEval-2010 Task 8 dataset show that the proposed model outperforms most state-of-the-art methods with an 86.2% F1-score.

Aiming at the low efficiency of A ^* algorithm in Digital Elevation Model (DEM) path planning, an improved A ^* algorithm based on distance and slope was proposed. A new evaluation function were designed by using distance and slope regarded as evaluation indexes in regular grid digital elevation model, and the pathability of surface barrier was judged. And in order to ensure that the improved algorithm was adaptive to the changing of the resolution ratio for DEM data, the parameters of the evaluation function were calculated according to the DEM data of the actual scene in the path searching process. Finally, a dynamic weight was changed with the changing of path searching, which could optimize path selection by adjusting the influence of completeness function and heuristic function on evaluation result. The simulation results show that the improved algorithm can adapt to the changing of DEM resolution by parameter adjustment, search the optimized path, reduce the search time and improve the search efficiency.

To solve the difficulty of attribute correlation in the process of simulating table data, an H model was proposed for describing not-temporal attribute correlation in table data. Firstly, the key attributes of the evaluation subject and the evaluated subject were extracted from the data set, by the twofold frequency statistics, four relationships of the key attributes were obtained. Then, the Maximum Information Coefficient (MIC) of each relationship was calculated to evaluate the correlation of each relationship, and each relationship was fitted by the Stretched Exponential (SE) distribution. Finally, the data scales of the evaluation subject and the evaluated subject were set. According to the result of fitting, the activity of the evaluation subject was calculated, and the popularity of the evaluated subject was calculated. H model was obtained through the association that was established by equal sum of activity and popularity. The experimental results show that H model can effectively describe the correlation characteristics of the non-temporal attributes in real data sets.

During the process of software development, activities of programmers including copy and paste result in a lot of code clones. However, the inconsistent code changes are always harmful to the programs. To solve this problem, and find harmful code clones in programs effectively, a method was proposed to predict harmful code clones by using Bayesian network. First, referring to correlation research on software defects prediction and clone evolution, two software metrics including static metrics and evolution metrics were proposed to characterize the features of clone codes. Then the prediction model was constructed by using core algorithm of Bayesian network. Finally, the probability of harmful code clones occurrence was predicted. Five different types of open-source software system containing 99 versions written in C languages were tested to evaluate the prediction model. The experimental results show that the proposed method can predict harmfulness for clones with better applicability and higher accuracy, and further reduce the threat of harmful code clones while improving software quality.

Aiming at the disadvantages of high fixing cost and partial observability of system environment in the process of repairing Autonomous Underwater Vehicle (AUV) software faults, a method was proposed based on micro-reboot mechanism and Partially Observable Markov Decision Process (POMDP) model for failure repair of AUV. To facilitate the implementation of the fine-grained self-repair micro-reboot strategy, a hierarchical structure was built based on micro-reboot combined with the characteristics of AUV software. Meanwhile, a self-repair model was put forward according to the theory of POMDP. With the goal of minimizing the fixing cost, the repair strategy was solved by Point Based Value Iteration (PBVI) algorithm to allow the repair action to execute in the partially observable environment at a lower cost.The simulation results show that the proposed repairing method can solve the AUV software failures caused by the software-aging and system calls. Compared with two-tier micro-repair strategy and three-tier micro-repair fixing strategy, this method is obviously superior to the contrast method in cumulative fault repair time and operational stability.

Conventional approaches for Common Weights (CW) generation in Data Envelopment Analysis (DEA) are either non-linear or scale-relevant. To solve this problem, according to the demand of military training performance evaluation, a new method was proposed to generate CW in DEA. The new method took DEA efficient units as the basis of calculation. Firstly, training data were normalized, and then multi-objective programing was employed for CW generation, which can lead to a fairer and more reasonable ranking of performances. The proposed method is not only linear, but also scale-irrelevant. Lastly, a military application illustrates that the proposed method is scientific and effective.

In order to solve the difficulty of selecting cluster centers and weakness of density analysis generalization, a novel clustering method was proposed. The method completed clustering by density and distance analysis based on genetic algorithm, which computed density with exponential method to reduce the impact of parameters and adopted genetic algorithm to search optimum threshold values. It introduced a penalty factor to overcome the excursion of search region for accelerating convergence. Numerical experiments on both artificial and UCI data sets show that compared with K-means, fast search clustering and Max_Min_SD, the proposed algorithm can achieve better or comparable performance on Rand index, accuracy, precision and recall.

To solve the practical problem that some rules may be lost when the association rules are extracted directly after the construction of interval concept lattice for the different formal context, the different interval concept lattices must be merged firstly. To improve the efficiency of lattice generating and consolidation, the incremental generation algorithm of interval concept lattice should be improved firstly, and then the concepts were stored in the form of structures which were divided into existence concepts, redundancy concepts and empty concepts. Secondly, the binary relation between extension and intension was analyzed and the sufficient condition of vertical merger, consistency of interval concept lattice, was defined. Thirdly the concepts which have consistent intension were divided into six kinds after merging and the corresponding decision theorem was given. In the end, based on the principle of breadth-first, a new vertical integration algorithm was designed through the type judgment and different processing methods of the concept lattice nodes in the original interval concept lattice. Finally, an application example verified the effectiveness and efficiency of the algorithm.

A novel kernel-based non-negative sparse representation (KNSR) method was presented for face recognition. The contributions were mainly three aspects: First, the non-negative constraints on representation coefficients were introduced into the Sparse Representation (SR) and the kernel function was exploited to depict non-linear relationships among different samples, based on which the corresponding objective function was proposed. Second, a multiplicative gradient descent method was proposed to solve the proposed objective function, which could achieve the global optimum value in theory. Finally, local binary feature and the Hamming kernel were used to model the non-linear relationships among face samples and therefore achieved robust face recognition. The experimental results on some challenging face databases demonstrate that the proposed algorithm has higher recognition rates in comparison with algorithms of Nearest Neighbor (NN), Support Vector Machine (SVM), Nearest Subspace (NS), SR and Collaborative Representation (CR), and achieves about 99% recognition rates on both YaleB and AR databases.

Considering the demand of detecting Android malware and the redundancy of permission properties, a fast scheme was proposed to detect malware from the perspective of permission correlation. To eliminate the redundant permissions, Chi-square test was used to compute the influence of the permission on the classification results. Then some representative permissions were selected on the basis of permission clustering to further reduce redundancy. Finally an improved Naive Bayesian classification based on the weights of different permissions was proposed to classify the software. Results of the experiments conducted on 2000 software samples show that the miss rate of malware detection is 10.33% and the overall prediction accuracy is 88.98%. Experiments indicate that this scheme is capable of detecting malware on Android platform by using a few permission properties, which can provide a reference for further analysis and judgment.

In order to forward messages more efficiently in the social oriented Delay Tolerant Network (DTN), this paper proposed a way to be aware of nodes' centralities. This paper came up with a routing mechanism for the awareness of nodes' centralities through a comprehensive analysis of the level of activity and the capacity of handling messages which aimed at quantifying the nodes' centralities. The result shows that compared with the traditional Epidemic routing and Contact Counts (CC) routing which principle is still based on nodes' centrality, the routing mechanism this paper discussed can significantly improve the rate of the message delivery and the rate of the overhead.

The exiting model representation cannot meet the requirements of multistage modeling, so model share, reuse and management can hardly achieve in multistage modeling process. Therefore, a hierarchical model management framework based on universal relation model presentation was presented. Firstly, the requirements of model representation in model management and the limitations of exiting model representation were analyzed, then a model representation based on universal relation was investigated to set the mapping relation between layers of conceptual model and mathematical model, and the integrative model representation and the hierarchical model management framework including conceptual model, mathematical model and physical model were set up. At last, the logic of modelbase in management framework was designed and the physical model generation based on universal relation was investigated. The model of different modeling process could be managed uniformly based on an integrative model representation.

In order to forecast the number of equipment maintenance people more easily and validly, a common approach of selecting the features of input vector in Support Vector Machine (SVM) named Main Factor Method (MFM) was proposed. The relevant terms of "main factor", "driving factor", "voluntary action" and "actions' carrier" were defined, based on which the theoretical MFM was constructed. Firstly, the predicting vector's main factor of voluntary actions was setup by "infinitely related principle" and "action purpose" method. Then the driving factors which can be looked as the characteristics of SVM input vector were refined through the selected main factor and "selecting principles of driving factors". The experimental results and comparison with other congeneric methods show that the proposed method can select the more accurate prediction with the value of relative average error 0.0109.

In this paper, a fuzzy clustering algorithm based on w-mean distance was proposed to solve such defects of Fuzzy C-Means (FCM) algorithm as easily falling into local optimal value and being sensitive to clustering center and noise data. First, initial clustering centers were determined by making use of the idea of the mean distance according to the distribution of data set, and the regulating factor w was introduced to adjust the mean distance. Second, each sample in data set was assigned a weight, and the clustering center formula and target function formula were modified by the weight, so that the anti-noise performance was greatly improved for the algorithm. In the end, the experimental results validate that the proposed algorithm has good effects on selecting initial clustering centers, avoiding local convergence, and having higher performance of anti-noise and effectiveness.

The effect data of actions show a significant randomness because of lots of uncertain elements in the course of action. In order to explore the rules of warfare hidden behind the data, the effectiveness evaluation was studied based on statistical analysis method. The basic concept of action and its effectiveness were analyzed. With the simulation data produced by enhanced irreducible semi-autonomous adaptive combat neural simulation toolkit (EINSTein), a single, a group and multi group experimental methods were used to study the statistical characteristics of offensive actions and find out that to a party who has a combat advantage, compared with increased number of personnel, the increased radius of firepower can achieve better operational results. On this basis, an evaluation method of action effectiveness was proposed and validated with simulation data. Therefore, a feasible resolution is provided to evaluate the action effectiveness based on actual combat data.

The application-layer multicast in end-system has overwhelming advantages compared with network-layer multicast. For improving the efficiency and performance of application-layer multicast, a multicast algorithm based on maximum interference network coding was presented. After adopting network coding, the new algorithm selected the maximum interference paths from source to every destination to improve coding efficiency and save network bandwidth. The simulations show that compared with non-coding multicast and simple network coding multicast algorithm, the new algorithm performs better in network throughout put and resource utilization.

Gaussian mixture background model is an online parameterized statistical model, and the presenting way of pixel sample pattern observed from time window has great influence on the model's learning result. According to the characteristics of dynamic background changes, issues such as the stability and plasticity of modal learning, the modal residual and activation, which affected the model's adaptive ability, were studied. The simulation results show that Gaussian mixture background model has a robust selective adaptability to gradual change, but a limited adaptability to transient variation of background configuration provided by modal residual and activation mechanism.

 Textile texture has much anomaly,because of yarn helix structure,different size of yarn and supple transform for textile itself. The methods for detecting the textile defects with feature and model to segment were low efficiency and not good enough in precision. A method of feature extraction of the textile defects by using Pulsed Coupled Neural Network(PCNN) was put forward to overcome these problems. The model and properties of PCNN was analyzed. According to different gray intensity between the field of textile defects and the field of normal textile, feature of textile defects were extracted for PCNN firing or not. After dilated, the textile defects’ edge were extracted with CANNY operator. Experiment shows that the method can much better get the feature of the textile defects and has a much better edge detection result of the textile defects.