As it is hard to recognize marine fishes occurred in submarine observation videos due to the bad undersea environment and low quality of the video, a recognition method based on deep learning was proposed. Firstly, the video was split into pictures, and as this type of video contains a large proportion of useless data, a background subtraction algorithm was used to filter the pictures without fish to save the time of processing all data. Then, considering the undersea environment is blurring with low bright, based on the dark channel prior algorithm, the pictures were preprocessed to improve their quality before recognition. Finally, a recognition deep learning model based on Convolutional Neural Network (CNN) was consructed with weighted convolution process to improve the robustness of the model. The experimental results show that, facing submarine observation video frames with poor quality, compared with traditional CNN, the method with preprocessing and weighted convolution as hidden layer can increase the recognition accuracy by 23%, contributing to the recognition of marine fishes in submarine observation video.

In the process of programming and source code reuse, since simple keyword-based code search often leads to inaccurate results, an accurate search method for source code was proposed. Firstly, according to the objectivity and uniqueness of syntax and semantics, the syntactic structure and semantics of I/O of a function in source code were considered as part of a query. Such query should be submitted following a regularized format. Secondly, the syntactic structure, semantics of I/O, keyword-compatible match algorithms along with the reliability calculation algorithm were designed. Finally, the accurate search method by combining syntactic and semantic queries was realized by using the above algorithms. The test result shows that the proposed method can improve Mean Reciprocal Rank (MRR) by more than 62% compared with the common keyword-based search method, and it is effective in improving the accuracy of source code search.

To address the low matching rate of flow table, an intelligent update method for flow table in Software Defined Network (SDN) switch was proposed. First, the impact of timeout value on the packet matching was described, besides, the shortcomings of First In First Out (FIFO), Least Recently Used (LRU) and other common methods were analyzed and compared. Secondly, based on the reality of survival time of the flow entry related closely to the characteristics of data flow, the Hidden Markov Model (HMM)-based Deep Flow Inspection (DFI) technology was used to classify the data flow. Finally, according to the condition of the flow table resources and controller's computing resources, the intelligent update of the flow entry of different type of data flow was realized. The simulation experiments conducted on data center behavior data of real campus indicate that the proposed method can improve more than 5% of the matching rate compared with the common methods, and it has a practical significance to the management of the SDN switch.

Concerning the problem of high energy consumption existing in Cloud storage Content Distribution Network (CCDN), the energy consumption optimization method for the CCDN was studied. Firstly, the operation principle of CCDN was analysed. Then, the energy consumption formulas were given for each cloud server and each network link. Moreover, the weighted graph was used to describe the whole network. Furthermore, based on the weighted graph, an energy consumption optimization algorithm named Min-Energy-Graph (MEG) was designed to satisfy the Quality of Service (QoS) of CCDN and data distribution of network system. MEG was compared with Greedy Site (GS) and Optimal Static Placement and Routing (OSPR) by the simulation experiments. Compared with the GS and OSPR, the energy consumption of MEG was reduced by 6.6% and 30% respectively in the experiment of system extension, the energy consumption of MEG was reduced by 28.9% and 60.2% separately in the experiment of ensuring the user QoS, and the energy consumption of MEG was reduced by 32.2% and 89.3% independently in the experiment of network topology density. The experimental results show that the proposed energy management method can greatly reduce the energy consumption of CCDN.

In wide-area Software Defined Networking (SDN) deployments, logically centralized controllers are usually physically placed in a distributed manner. To address the unbalanced controller load problem in controller placement, a load-aware optimal placement for heterogeneous controllers based on unbalanced graph partitioning was proposed. First, the controller placement problem and related metrics, controller load balance and latency, were analyzed. Secondly, controller load balance and latency caused by controller placement were quantified and computed using graph theory and cosine similarity. Moreover, based on graph partitioning theory, the load-aware controller placement problem was transformed into a specified graph partitioning problem. Finally, a load-aware optimal controller placement scheme was presented based on multi-level graph partitioning. Simulation experiments conducted on real network topologies indicate that the proposed scheme can effectively achieve nearly optimal controller load balance.

Concerning the problem that previous studies mostly consider from the resource provider's perspective, and user's evaluations have not been fully utilized to improve the resource decision making ability, this paper proposed a resource re-allocation method focusing on the user's evaluation feedback. First, through analyzing the process of cloud resource allocation, several factors influencing decision-making were defined, and an adaptive cloud resource management framework with user's involvement was proposed. Next, the main idea of method of resource re-allocation with user's involvement was elaborated, and a formula was designed to guide user's evaluation. Finally, based on similarity theory, the user's expected satisfaction of a new cloud task was predicted. Together with the cloud task parameters and environment parameters, it was used to be the input of BP (Back Propagation) neural network to make the resource allocation decision. In the comparison experiments with the allocation scheme without user's involvement, the average user's satisfactory of the proposed scheme increased by 7.4%, maintained at more than 0.8, showed a steady upward trend. In the comparison experiments with Min-Max algorithm and Cloud Tasks-Resources Satisfactory Matching (CTRSM) algorithm, its average user's satisfactory increased by 16.7% and 4.6% respectively. The theoretical analysis and simulation results show that the cloud resource re-allocation method focusing on user's evaluation is self-improved, and it can improve the adaptive ability of cloud resource management.

The problem of cloud computing processors failure cannot be ignored in the cloud environment. Fault-tolerance becomes a key requirement in the design and development of cloud computing systems. Aiming at the problem of low scheduling efficiency and single type of task in most fault-tolerant scheduling algorithms, the fault-tolerant scheduling method based on processors, primary-backup copies of hybrid tasks grouped was proposed. A method to determine whether two backup copies can overlap was presented. What's more, the calculation formulas of periodic task worst-case response time and completion time of aperiodic tasks preemptive execution were given. The simulation result shows that the proposed algorithm has a remarkable saving of cloud computing system processors needed and scheduling computation time compared with Hybrid real time task Fault Tolerant Scheduling (HFTS) algorithm. It is of great significance for improving the reliability of cloud system and the schedulability of real-time tasks set, as well as the processor efficiency.

The number of components is continuously growing in the network component library, it is hard for users to select high-quality components from the massive uneven-quality components. In order to solve the problem, a reuse trust evaluation method based on component use dependency relations was proposed, in which component base was used as an evidence base. Firstly, component dependency relations were collected from evidence base. Secondly, the basic trust function was defined for each component, and the different believable weight value was set up for each evidence according to the different sources of component dependency relations on the above basis. Finally, the final trust value of component was generated by a specific conversion algorithm with the obtained results. In instance analysis, the component evaluation result of the proposed method was consistent with the expectation and the conclusion which was gotten by the internal and external quality model of the reference components. However, the proposed method greatly reduced the workload of component's credible evaluation and improved the evaluation efficiency. The results of analysis show that the proposed method can objectively reflect the credibility of components, and can be used as a trusted measurement mechanism of component retrieval in the component library, which helps to realize the high quality retrieval and reuse of components.

In E-commerce website, massive disorder shopping reviews may make the consumers be lost in the massive shopping reviews and can not distinguish trusted reviews. Therefore, this paper proposed a trustworthy sort method for customer reviews. Firstly, focusing on commercial advertising information in websites and concerning about whether the contents of the online customer reviews and product functional properties are closely related, the authors designed an algorithm of product's key features extractions from shopping websites based on HTML script format, and presented a method of customer reviews features extractions based on natural language processing. Secondly, the authors used the technique of words similarity to analyze the correlation degree between product features and customer reviews contents, and then proposed the computational method of trust degree for shopping customer reviews. Finally, through analyzing the method with an example, the proposed method achieves a trustworthy sort for large online shopping customer reviews. Thus customers need not browse all reviews to judge which one can be trusted or have the real reference value. It decreases information search costs and improves the efficiency of decision making.

Concerning the problem that previous studies on the scalability do not fully consider parallel execution time, and the relationships between latency scalability and parallel execution time have not been yet studied thoroughly, this paper studied the relationships between latency scalability and parallel execution time deeply and fully. Thereby some important conclusions were drawn, and they were about the relationships between latency scalability and parallel execution time after different algorithm-machines were extended from the same initial state. Then the proof of the above conclusions was given in this paper. The derived conclusions enriched the research content about the relationships between latency scalability and parallel execution time and provided a theoretical basis for obtaining ideal latency scalability of parallel computing. Finally the important conclusions and analytical expressions were verified through experimental results obtained for different algorithm-machines.

It is more and more difficult to find the valuable required scientific papers accurately and efficiently on Internet, thus a new thesis evaluation method was proposed based on consistency of the title and text to deal with this problem. First of all, the title and text were modeled by eigenvectors respectively. After that, the technique of words similarity was used to calculate the matching-degree of each feature word in title and text vector. The feature word pair was successfully matched if their matching degree was greater than a certain threshold. Then all such matching pairs and their word weights were counted up to calculate the credibility of the title. Based on the hierarchical tree structure of the thesis title, the similarity matching degree of all headings and their corresponding text were calculated by Depth First Traversal (DFT) algorithm, and then the credibility of the paper was evaluated. A case study results prove that the proposed method can realize the scientific papers' credible quality assessment, which makes it be more efficient for readers in paper reading.

In practical engineering applications, using traditional CPU serial computation for combustion numerical simulation can hardly meet the requirements on simulation speed. This paper took the advantage of GPU which has more computing capability than CPU, by discretizing the combustion physical equations on staggered grid, solving the discrete equations with preconditioned bi-conjugate gradient stabilized (PBiCGSTAB) method, exploring the parallel algorithm of GPU-oriented matrix vector multiplication and the parallel algorithm of inverse matrix vector multiplication. Hence a feasible method for numerical calculation of laminar diffusion combustion on GPU was offered. The experimental results indicate that the parallel program on GPU achieved more than 10 times speedup relative to serial program on CPU. Since the calculated results are in line with the actual situation, the method is feasible and efficient.

This paper proposed a new dynamic software watermarking algorithm based on stack-state relations in order to overcome the weakness of the existing software watermarking algorithms. The watermark was hidden in the stack-state relations generated by execution of the program and extracted by recognizing the relationship of stack states in runtime. The paper gave the concept of stack-state transition graph, and explained its properties. It also gave detailed description of the software watermark embedding and extraction algorithms, and explained it through a specific example. The analysis and experiments show that the algorithm does not affect the host program significantly, and has the ability to resist various attacks.