Since Practical Byzantine Fault Tolerance （PBFT） consensus algorithm applied to consortium blockchain has the problems of insufficient scalability and high communication overhead， an improved practical Byzantine fault tolerance consensus algorithm based on Raft algorithm named K-RPBFT （K-medoids Raft based Practical Byzantine Fault Tolerance） was proposed. Firstly， blockchain was sharded based on K-medoids clustering algorithm， all nodes were divided into multiple node clusters and each node cluster constituted to a single shard， so that global consensus was improved to hierarchical multi-center consensus. Secondly， the consus between the cluster central nodes of each shard was performed by adopting PBFT algorithm， and the improved Raft algorithm based on supervision nodes was used for intra-shard consensus. The supervision mechanism in each shard gave a certain ability of Byzantine fault tolerance to Raft algorithm and improved the security of the algorithm. Experimental analysis shows that compared with PBFT algorithm， K-RPBFT algorithm greatly reduces the communication overhead and consensus latency， improves the consensus efficiency and throughput while having Byzantine fault tolerance ability， and has good scalability and dynamics， so that the consortium blockchain can be applied to a wider range of fields.

In consideration of the problems of low utilization rate of action information and insufficient attention of temporal information in video human action recognition, a human action recognition model based on tightly coupled spatiotemporal two-stream convolutional neural network was proposed. Firstly, two 2D convolutional neural networks were used to separately extract the spatial and temporal features in the video. Then, the forget gate module in the Long Short-Term Memory (LSTM) network was used to establish the feature-level tightly coupled connections between different sampled segments to achieve the transfer of information flow. After that, the Bi-directional Long Short-Term Memory (Bi-LSTM) network was used to evaluate the importance of each sampled segment and assign adaptive weight to it. Finally, the spatiotemporal two-stream features were combined to complete the human action recognition. The accuracy rates of this model on the datasets UCF101 and HMDB51 selected for the experiment and verification were 94.2% and 70.1% respectively. Experimental results show that the proposed model can effectively improve the utilization rate of temporal information and the ability of overall action representation, thus significantly improving the accuracy of human action recognition.

Facial action unit detection aims at making computers detect the action unit targets based on the given facial images or videos automatically. Due to a great amount of research during the past 20 years, especially the construction of more and more facial action unit databases and the raise of deep learning based methods, facial action unit detection technology has been rapidly developed. Firstly, the concept of facial action unit and commonly used facial action unit databases were introduced, and the traditional methods including steps such as pre-processing, feature extraction and classifier learning were summarized. Then, for several important research areas, such as region learning, facial action unit correlation learning and weak supervised learning, systematic review and analysis were conducted. Finally, the shortcomings of the existing reasearch and potential developing trends of facial action unit detection were discussed.

Since Practical Byzantine Fault Tolerance (PBFT) consensus algorithm applied to the alliance chain has the problems of static network structure, random selection of master node and large communication overhead, an Evolution of Practical Byzantine Fault Tolerance (EPBFT) consensus algorithm was proposed. Firstly, a series of activity states were set for consensus nodes, making the nodes have complete life cycle in the system through state transition, so that the nodes were able to dynamically join and exit while the system has a dynamic network structure. Secondly, the selection method of master node of PBFT was improved with adding the election process of master node with the longest chain as the election principle. After the election of master node, the reliability of master node was further ensured through data synchronization and master node verification process. Finally, the consensus process of PBFT algorithm was optimized to improve the consensus efficiency, thus the communication overhead of EPBFT algorithm was reduced to 1/2 of that of PBFT algorithm with little view changes. The experimental results show that EPBFT algorithm has good effectiveness and practicability.

Aiming at the problem that the discontinuity of the pitch sequence of the same sound source was caused by the interference of different sound sources in polyphonic music which reduced the accuracy of pitch estimation, an improved pitch contour creation and selection algorithm for melody extraction was proposed. Firstly, a method based on auditory streaming cues and the continuity of pitch salience was proposed to create pitch contour by calculating the pitch salience of each point in the time-frequency spectrum. In order to further select the melody pitch contour, the non-melodic pitch contours were removed according to the repetitive characteristics of the accompaniment, and dynamic time warping algorithm was used to calculate the similarity between the melodic and non-melodic pitch contours. Finally, the octave errors in the melodic pitch contours was detected based on the long term relationship of the adjacent pitch contours. Simulation experiments on the data set ORCHSET show that the pitch estimation accuracy and the overall accuracy of the proposed algorithm are improved by 2.86% and 3.32% respectively compared with the oringinal algorithm, which can effectively solve the pitch estimation problem.

Extrapolation technique of weather radar echo possesses a widely application prospects in short-term nowcast. The traditional methods of radar echo extrapolation are difficult to obtain long limitation period and have low utilization rate of radar data. This problem is researched from deep learning perspective in this paper, and a new model named Dynamic Convolutional Neural Network based on Input (DCNN-I) was proposed. According to the strong correlation between weather radar echo images at adjacent times, dynamic sub-network and probability prediction layer were added, and a function was created that maped the convolution kernels to the input, through which the convolution kernels could be updated based on the input weather radar echo images during the testing. In the experiments of radar data from Nanjing, Hangzhuo and Xiamen, this method achieved higher accuracy of prediction images compared with traditional methods, and extended the limitation period of exploration effectively.

Focusing on the shortcoming of low storage efficiency and high recovery cost after copy failure of the current Hadoop, Hopfield Neural Network (HNN) was used to improve the overall performance. Firstly, the resource characteristics that affect the storage efficiency were analyzed. Secondly, the resource constraint model was established, the Hopfield energy function was designed and simplified. Finally, the average utilization rate of 8 nodes was analyzed by using the standard test case Wordcount, and the performance and resource utilization of the proposed strategy were compared with three typical algorithms including dynamic resource allocation algorithm, energy-efficient algorithm and Hadoop default storage strategy, and the comparison results showed that the average efficiency of the storage strategy based on HNN was promoted by 15.63%, 32.92% and 55.92% respectively. The results indicate that the proposed algorithm can realize the resource load balancing, help to improve the storage capacity of Hadoop, and speed up the retrieval.

Cloud task scheduling is a hot issue in the research of cloud computing. The cloud task scheduling method directly affects the overall performance of the cloud platform. A task scheduling method Template-Based Genetic Algorithm (TBGA) was proposed. Firstly, according to the processor's CPU speed, bandwidth and etc., the amount of tasks that should be allocated to each processor was calculated. andwas called allocation template. Secondly, according to the template, the tasks were combined into multiple subsets and finally each subset of tasks was allocated to the corresponding processor by using genetic algorithm. Experimental results show that the method can obtain shorter time scheduling for total tasks. TBGA reduced 20% of task set completion time compared with Min-Min algorithm and 30% of task set completion time compared with Genetic Algorithm (GA). Therefore, the TBGA is an effective task scheduling algorithm.

Concerning the problem that the existing Voice Activity Detection (VAD) algorithms based on Hidden Markov Model (HMM) were poor to track noise, a method using Baum-Welch algorithm was proposed to train the noise with different characteristics, and the corresponding noise model was generated to establish a library. When voice activity was detected, depending on the measured background noise of the speech, the voice was dynamically matched to a noise model in the library. Meanwhile, in order to meet real-time requirements of speech signal processing, reduce the complexity of the speech parameter extraction, the threshold was improved to ensure the inter-frame correlation of the speech signal. Under different noise environments, the improved algorithm performance was tested and compared with Adaptive Multi-Rate (AMR), G.729B of the International Telecommunications Union (ITU-T). The test results show that the improved algorithm can effectively improve the accuracy of detection and noise tracking ability in real-time voice signal processing.

To solve the problem of low resource utilization in community health center, little contact between community health center and community residents, and difficulty for residents to participate in personal health management and medical care, an intelligent healthy community system was developed. With the increasing popular mobile devices, the system provided support for health record management, chronic disease management, immunization, appointment registration, medical information query and other services in community health center. It realized the data sharing and interaction among smart phones, tablet PCs and Hospital Information System (HIS), which allowed the residents to actively participate in personal health management. Now the system has been deployed in one community health center of Chengdu, it makes community residents convient to manage their personal health, and improves the work efficiency and service quality of community health center.

Concerning that most Web video categorization researches just focus on the basic simple fusion of the information from text model and visual model, a Web video classification method based on the bidirectional propagation of heterogeneous attributes was proposed. Firstly, the method adopted K-means clustering to divide key frames into multiple clusters, and modeled videos at the level of frame. For each cluster, a part of key frames were randomly chosen to propagate their text information to the cluster. For each key frame, the text explanation of the corresponding cluster was transferred to this frame. Finally, the Web video was classified based on the extended text information from dual propagation by using Support Vector Machine (SVM) classifiers. The method integrates heterogeneous attributes well based on the dual propagation. The experimental results demonstrate the effectiveness of the method in the Web video classification.

To improve Dalvik virtual machine performance, a hybrid concurrent model based on multi-thread mechanism for Java virtual machine was proposed. This model implemented performance optimization by overlapping the production of native code with program execution through multi-thread control and hot-function table, and some critical issues in the design and implementation were also discussed. The experimental results show that this model can improve the Java execution speed efficiently in Dalvik.

Cross-layer design is an effective technique for future mobile communication systems. A cross-layer resource allocation algorithm with partial channel state information was explored to maximize the total system throughput for multi-user MIMO-OFDM （Multiple Input Multiple Output-Orthogonal Frequency Division Multiplexing） system. The objective function of the optimization problem was designed based on the power limitation constraint, transmission rate, average queue length and sub-carrier occupancy, Quality of Service (QoS) requirements of different services and queue state information of data link layer. Under the condition of finite-length user buffer in data link layer, the mean feedback model was utilized to describe the feedback process of channel state information, and then the corresponding cross-layer resource allocation criteria could be derived. The simulation results show that compared with the existing schemes, the proposed algorithms obtain reasonable throughput performance and reduce lost package rate while providing better QoS requirement for each user of different services.

Clustering can be used in the routing algorithm to enhance the scalability of Wireless Sensor Network (WSN). Concerning the defects of traditional clustering algorithm, a new strategy EDC (Energy-efficient, Dual-path, Clustering) was proposed, in which the member node has an optimal backup path. The strategy guaranteed that member node can still transmit data through its backup path when its cluster head was dying in the WSN. The results of the simulation experiment on the platform OMNeT ++ indicate that EDC performs much better than other protocols of WSN in terms of network reconstruction time and number of failed nodes.

To improve the quality of video surveillance outdoors and to automatically acquire the weather situations, a method to recognize weather situations in outdoor images is presented. It extracted such parameters as power spectrum slope, contrast, noise, saturation as features to realize the multi-classification of weather situations with Support Vector Machine (SVM). Then a decision tree was constructed in accordance with the distance between these features. The experimental results on WILD image base and our image set of eight hundred samples show that the proposed method can recognize sunny, overcast, foggy weather more than 85%, and recognize rainy weather more than 75%.

To increase the classification accuracy of the database classification system, this paper proposed a new classification method. Firstly, the continuous attributes were dispersed by the Fuzzy C-Mean (FCM) algorithm. Secondly, an improved fuzzy association method was proposed to mine the classification association rules. Eventually, the compatibility between the generated rules and patterns was used to construct a set of feature vectors, which were used to generate a classifier. The experimental results demonstrate that the method has high discrimination and efficiency.

Based on the current research on improving realtime PPM algorithms in IP traceback, the relations among marking probability, traffic volume for constructing an attack path and the distance of the attacking path were analysed. And an approach of realtime IP tracebacking which deploys larger traffic first probabilistic packet marking scheme (LTFMS) was proposed. According to the statistics on the present traffic of routers, a victim could construct a major attacking path in minimum time. For large-scale DDoS attacks, by establishing a simulated test environment and experiment analysis, LTFMS can construct more attacking paths than the existing schemes within the same time.

An embedded color facial image coding,based on the characteristics of facial image and multi-correlations of color image with wavelet decomposition, was proposed for low bit rate transmission. This algorithm combined joint-component vector quantization respectively with zero block and zero tree coding,scanned three component of color image separately and simplified separate coding into joint scanning and joint encoding. Experiment results show that in low bit rate environment, the new image compression scheme performs better than JPEG in the aspects of perception and PSNR(Peak Signal Noise Ratio).