Recently, botnets, domain name hijacking, phishing websites and other typosquatting attacks are more and more frequent, seriously threatening the security of society and individuals. Therefore, the typosquatting detection is an important part of network protection. The current typosquatting detections mainly focus on public domain names, and the detection methods are mainly based on edit distance which is difficult to fully reflect the visual characteristics of domain names. In addition, using the related information of the given domains for determination can help to increase the detection efficiency, but it also introduces a large additional cost. Based on this, a lightweight detection strategy only based on domain name strings was adopted for typosquatting detection. By comprehensively considering the influence of character locations, character similarities and operation types on the vision of domain names, the edit distance algorithm based on visual characteristics was proposed. According to the characteristics of typosquatting, firstly the domain names were preprocessed, then different weights were given to the characters according to their positions, character similarities and operation types, and finally, the typosquatting determination was performed by calculating the edit distance value. Experimental results show that compared with the detection method based on edit distance, the typosquatting lightweight detection method based on visual features has the F1 value increased by 5.98% and 13.56% respectively when the threshold value is 1 and 2, which proves that the proposed method has a good detection effect.

In order to solve the problem of insufficient available training data in the classification task of breast mass and calcification, a multi-view model based on secondary transfer learning was proposed combining with imaging characteristics of mammogram. Firstly, CBIS-DDSM (Curated Breast Imaging Subset of Digital Database for Screening Mammography) was used to construct the breast local tissue section dataset for the pre-training of the backbone network, and the domain adaptation learning of the backbone network was completed, so the backbone network had the essential ability of capturing pathological features. Then, the backbone network was secondarily transferred to the multi-view model and was fine-tuned based on the dataset of Mianyang Central Hospital. At the same time, the number of positive samples in the training was increased by CBIS-DDSM to improve the generalization ability of the network. The experimental results show that the domain adaption learning and data augmentation strategy improves the performance criteria by 17% averagely and achieves 94% and 90% AUC (Area Under Curve) values for mass and calcification respectively.

There are two main problems about pneumothorax detection in X-ray images. The pneumothorax usually overlaps with tissues such as ribs and clavicles in X-ray images, easily causing missed diagnosis and the performance of the existing pneumothorax detection methods remain to be improved. The suspicious pneumothorax area detection cannot be exploited by the convolutional neural network-based algorithms, lacking the interpretability. Aiming at the problems, a novel method combining Dense convolutional Network (DenseNet) and gradient-weighted class activation mapping was proposed. Firstly, a large-scale chest X-ray dataset named PX-ray was constructed for model training and testing. Secondly, the output node of the DenseNet was modified and a sigmoid function was added after the fully connected layer to classify the chest X-ray images. In the training process, the weight of cross entropy loss function was set to alleviate the problem of data imbalance and improve the accuracy of the model. Finally, the parameters of the last convolutional layer of the network and the corresponding gradients were extracted, and the areas of the pneumothorax type were roughly located by gradient-weighted class activation mapping. The experimental results show that, the proposed method has the detection accuracy of 95.45%, and has the indicators such as Area Under Curve (AUC), sensitivity, specificity all higher than 0.9, performs the classic algorithms of VGG19, GoogLeNet and ResNet, and realizes the visualization of pneumothorax area.

In order to effectively protect video information, according to the characteristics of H.265/HEVC (High Efficiency Video Coding), a scheme combining transform coefficient scrambling and syntax element encryption was proposed. For Transform Unit (TU), the TU with the size of 4×4 was scrambled by Arnold transform. At the same time, a shift cipher was designed, and the cipher was initialized according to the approximate distribution rule of the Direct Current (DC) coefficient of the TU, and the DC coefficients of TU with the size of 8×8, 16×16 and 32×32 were shifting encrypted using encryption map generated by Arnold transform. For some of the syntax elements with bypass coding used in the entropy coding process, the Logistic chaotic sequence was used for encryption. After encryption, the Peak Signal-to-Noise Ratio (PSNR) and Structual Similarity (SSIM) of the video were decreased by 26.1 dB and 0.51 respectively on average, while the compression ratio was only decreased by 1.126% and the coding time was only increased by 0.17%. Experimental results show that under the premise of ensuring better encryption effect and less impact on bit rate, the proposed scheme has less extra coding overhead and is suitable for real-time video applications.

In recent years, the tendency of depression tends to occur at a younger age and affects more people. Although research on the topic has achieved some results, it still lacks a more objective and accurate method for identifying depressive tendencies, and research on depressive tendencies from multiple perspectives is lacking. Therefore, the combination of mental health self-check table and eye-tracking was proposed as a method for identifying depressive tendencies and was studied from multiple perspectives. The innovative features of eye movement, memory, cognitive style, and network behaviors were incorporated. In order to address complex feature relationship and extract more useful information, a scanning process with combining a stacking method was proposed to form a proposed recognition model for depressive tendencies called scanning stacking model. To comprehensively and objectively evaluate the performance of scanning and stacking model, the independent contributions of both scanning process and stacking method were evaluated in the experiment. The experimental results show that the independent contribution of scanning process is 0.03, and the independent contribution of stacking method is 0.02. In addition, the scanning stacking model was compared with several models from parameter R-squared, Mean Square Error (MSE) and average absolute error, and the results show that the scanning stacking model has better prediction effect.

Although the relevant research on Web advertisement effect has achieved good results, there are still a lack of thorough research on the interaction between advertisement and each blue link in a Web page, as well as a lack of thorough analysis of the impact of user characteristics and advertising features, and advertising metrics are also inappropriate. Therefore, a method based on multi-modal feature fusion was proposed to study the effectiveness of Internet advertising and user behavior patterns under the influence of multiple factors. Through the quantitative analysis of multi-modal features, the attractiveness of advertising was verified, and the attention effects under different conditions were summarized. By mining frequent patterns of user behavior information and combining with the characteristics of the data, the Directional Frequent Browsing Patterns (DFBP) algorithm was proposed to directionally mine the most common browsing patterns of users with fixed-length. Memory was used as a new index to measure the quality of advertising, and the random forest algorithm was improved by frequent pattern, then a new advertising memory model was built by fusing multimodal features. Experimental results show that the memory model has an accuracy of 91.64%, and it has good robustness.

Ultrasound imaging is one of the most important diagnostic techniques of modern medical imaging. However, due to the presence of multiplicative speckle noise, the development of ultrasound imaging has been limited. For this problem, an improved strategy for Bayesian Non-Local Means (NLM) filtering algorithm was proposed. Firstly,a Bayesian formulation was applied to derive an NLM filter adapted to a relevant ultrasound noise model, which leads to two methods of calculating distance between the image blocks, the Pearson distance and the root distance. Secondly, to lighten the computational burden, a image block pre-selection process was used to accelerate the algorithm when a similar image block was selected in the non-local area. In addition, the relationship between parameter and noise variance was determined experimentally, which made the parameter being adaptive to the noise.Finally, the VS (Visual Studio) and OpenCV (Open source Computer Visual library) were used to realize the algorithm, making the program running time greatly reduced. In order to evaluate the denoising performance of the proposed algorithm, experiments were conducted on both phantom images and real ultrasound images. The experimental results show that the algorithm has a great improvement in the performance of removing speckle noise and achieves satisfactory results in terms of preserving the edges and image details, compared with some existing classical algorithms.

The SQL (Structured Query Language) injection attack is a threat to Web applications. Aiming at SQL injection behaviors in PHP (Hypertext Preprocessor) applications, a model of detecting SQL injection behaviors based on tainting technology was proposed. Firstly, an SQL statement was obtained when an SQL function was executed, and the identity information of the attacker was recorded through PHP extension technology. Based on the above information, the request log was generated and used as the analysis source. Secondly, the SQL parsing process with taint marking was achieved based on SQL grammar analysis and abstract syntax tree. By using tainting technology, multiple features which reflected SQL injection behaviors were extracted. Finally, the random forest algorithm was used to identify malicious SQL requests. The experimental results indicate that the proposed model gets a high accuracy of 96.9%, which is 7.2 percentage points higher than that of regular matching detection technology. The information acquisition module of the proposed model can be loaded in an extended form in any PHP application; therefore, it is transplantable and applicable in security audit and attack traceability.

Focusing on the problems of non-adaption of filtering parameters and edge blur of Non-Local Means (NLM) algorithm, an improved NLM denoising algorithm based on image segmentation was proposed. The proposed algorithm is composed of two phases. In the first phase, the filtering parameter was determined according to the noise level and image structure, and traditional NLM algorithm was used to remove the noise and generate the rough clean image. In the second phase, the estimated clean image was divided into detailed region and background region based on pixel variance, and the image patches belonged to different regions were denoised separately. To effectively remove the noise, the back projection was utilized to make full use of the residual structure from the method noise of the first phase. The experimental results show that compared with traditional NLM and three NLM-improved algorithms, the proposed algorithm achieves higher Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM), while maintaining more structure details and edges, making the denoised image clear and retaining the complete real information.

Since the existing block-sparse system identification algorithm based on Mean Square Error (MSE) shows poor performance under impulsive interference, an Improved Block Sparse-Normalization Least Mean Square (IBS-NLMS) algorithm was proposed by introducing the inverse hyperbolic sine cost function instead of MSE. A new cost function was constructed and the additive value was obtained by steepest-descent method. Furthermore, a new vector updating equation for filter coefficients was deduced. The adaptive update of the weight vector was close to zero in the presence of impulsive interference, which eliminated the estimation error of adaptive updating based on the wrong information. Meanwhile, mean convergence behavior was analyzed theoretically and then the simulation results demonstrate that in comparison with the Block Sparse-Normalization Least Mean Square (BS-NLMS) algorithm, the proposed algorithm has higher convergence rate and less steady-state error under non-Gaussion noise impulsive interference and abrupt change.

For the problem of denoising images corrupted by mixed noise such as Additive White Gaussian Noise (AWGN) with Salt-and-Pepper Impulse Noise (SPIN) and Random-Valued Impulse Noise (RVIN), an improved image restoration algorithm on the basis of the existing weighted encoding method was proposed. The image priors about sparse representation and non-local similarity were integrated. Firstly, the sparse representation based on the dictionary was used to build a variational denoising model and a weighting factor was designed for data fidelity term to suppress impulse noise. Secondly, the method of non-local means was used to get an initialized denoised image and then a mask matrix was built to remove impulse noise points to get the good non-local similarity prior knowledge. Finally, the image sparsity prior and non-local similarity prior were integrated into the regularization of the variational model. The final denoised image was obtained by solving the variational model. The experimental results show that in different noise ratios, the Peak Signal-to-Noise Ratio (PSNR) of the proposed algorithm increased 1.7 dB than that of fuzzy weighted non-local means filter, and the Feature Similarity Index (FSIM) increased 0.06. Compared with weighted encoding method, the PSNR increased 0.64 dB, and the FSIM increased 0.03. The proposed method has better recovery performance especially for the texture strong images and can retain real information of the image.

In order to solve the poor robust problem of Automatic Speech Recognition (ASR) system in noisy environment, a new estimation algorithm of switching speech power spectrum was proposed. Firstly, based on the assumption of the speech spectral amplitude was better modelled for a Chi distribution, a modified estimation algorithm of speech power spectrum based on Minimum Mean Square Error (MMSE) was proposed. Then incorporating the Speech Presence Probability (SPP), a new MMSE estimator based on SPP was obtained. Next, the new approach and the conventional Wiener filter were combined to develop a switch algorithm. With the heavy noise environment, the modified MMSE estimator was used to estimate the clean speech power spectrum; otherwise, the Wiener filter was employed to reduce calculating amount. The final estimation algorithm of switching speech power spectrum for ASR system was obtained. The experimental results show that,compared with the traditional MMSE estimator with Rayleigh prior, the recognition accurate of the proposed algorithm was averagely improved by 8 percentage points in various noise environments. The proposed algorithm can improve the robustness of the ASR system by removing the noise, and reduce the computational cost.

The parallel Particle Swarm Optimization (PSO) algorithm was improved through Graphics Processor Unit (GPU) based on Compute Unified Device Architecture (CUDA). According to the structural characteristics of the CUDA hardware system, it can be concluded that block is executed serially and the basic scheduled and executive unit of Streaming Multiprocessor (SM) is warp. GPU parallel PSO algorithm based on adaptive warp was carried out in order to make full use of thread parallelism in the block. The dimensions of particles were corresponded to the threads of particles. Each particle was corresponded to one or more warps in accordance with its self-dimension adaptively by using the warp level parallelism of GPU. One or more particles were corresponded to each block. Comparison with the existing coarse-grained parallel approach (corresponding each particle to the thread) and fine-grained parallel approach (corresponding each particle to the block) was made, and the experimental results show that the proposed parallel approach achieves CPU speed-up ratio of 40 more than two kinds of approaches mentioned above.

An image positioning and scaling architecture for mobile video terminals was proposed for freely zooming and viewing video in detail. Then a gesture recognition processing approach was adopted in the architecture. Single-finger dragging and double-finger zooming detection were proposed for the gesture recognition. In addition, an approach to coordinates conversion calculation was proposed with boundary binding of coordinate transformation parameters using crossing boundary detection. Novel video display system was presented which consists of the video decoding, the image rendering and the interaction with synchronization. Finally these parts were concurrently implemented by three threads. The simulation results show that the proposed system obtains real-time image positioning and scaling while the traditional way of video playback is reserved. Interaction response time is controlled within 6ms to eliminate the screen flicker and skipping caused by interaction. Real-time image positioning and scaling of video playback for resources-limited mobile terminals will lead to a wide range of potential applications.

The existing methods for backbone networks only analyze coarse-grained network traffic characteristic parameters. Thus, they cannot guarantee both the premise of abnormal flow identification and the real-time detection for DoS (Denial of Service) & DDoS (Distributed Denial of Service, DDoS) attacks. Concerning this problem, a DoSⅅoS attack detection and abnormal flow identification method for backbone networks was proposed. First, it analyzed coarse-grained network traffic characteristic parameters to determine the time points that abnormal behaviors occur; then, fine-grained traffic behavior characteristic parameters were analyzed in these time points to find the destination IP addresses that correspond to abnormal behaviors; finally, comprehensive analysis was conducted for extracted traffic that correspond to abnormal behaviors to determine DoS and DDoS attacks. The simulation results show that, the proposed method can effectively detect DoS attacks and DDoS attacks in backbone networks. Meanwhile, it could accurately identify the abnormal traffic, while real-time detection is ensured.

In order to overcome the deficiency of the weighted average strategy which is adopted in the structure similarity algorithm for the perception of image quality, considering that certain regions in an image may not bear the same importance as others, an image quality assessment metric based on local invariant features was put forward. The algorithm used structural similarity to calculate the quality map of distorted image, and then extracted the local invariant features points in the distorted image. The region around features points was given more visual importance, and the quality of the image distortion could be evaluated by using integrated weighting strategy. The experimental results on the standard image library show that the computational complexity of this algorithm is relatively lower and the evaluation performance of structure similarity algorithm can be considerably increased, which achieves better consistency with the subjective assessment of human eyes.

Aiming at the all-pairs shortest-path problem in the directed graph,a practical parallel algorithm,which based on the Floyd algorithm with an extended path array,was brought forward on 2-D mesh network.The planar evenly partition method was chosen for task division in this parallel algorithm.The parallel algorithm was implemented on MPI on NOW.The theoretical analysis and the experimental results prove that the parallel algorithm is an efficient and scalable algorithm.