Parallel Computing Methods And System For Image Retrieval

With the rapid development of computer technology, multimedia technology and communication technology, multimedia information is rapidly becoming the mainstream of the exchange of information and services. The Content-based image retrieval system has important application value, but its computational efficiency can not meet the processing requirements of massive network images. Parallel computing is one of the effective ways to solve the problem.The key technologies of Content-based image retrieval system for parallel computing are researched. Through the parallel computing between tasks and between parallel units within task, the purpose of high efficiency and accurate image retrieval is promised to be achieved. Around this purpose, the main research contents are data-driven task balancing mechanism and load balancing scheduling method. Multiple tasks are reasonably distributed by the task scheduling mechanism to different computer nodes, making the tasks executed in parallel as soon as possible to reduce the total execution time and to improve the running efficiency of the system. The load balancing scheduling is to reload jobs from the heavy load nodes to the light load nodes, so that all nodes have the equal load.The system periodically detects node status changes, with which, system dynamically adjusts the nodes task running. Introduced the server node load redundancy value, the current surplus computing power of each node as well as system total remaining computing resources can be effectively predicted, and over load of nodes as well as inconsistencies in data processing between tasks are effectively avoided. A kind of data-driven task balancing mechanism for redundant load is presented base on depth study of task scheduling. Its distribution is based on task balance level, and its measure is the average speed of processed images in a scheduling cycle time. Experimental results show that the proposed task balancing algorithm is effective. Combined with analysis and comparison of current load balancing algorithms, a fusion of multiple resources for dynamic load balancing method with adaptive scheduling is proposed. Its detailed process is that the changes of system nodes load and system computing resources are periodically detected, according to each node load state and system calculation ability, and then equals each node load dynamically. Experimental result shows that load of each node tends to be balanced. On this basis, a content-based image retrieval system for parallel computing with the hardware adaptive ability has being implemented. Finally, experimental results of the system are given, and the results show that processed 10000 images, the speedup is 1.30 by a computer parallel and speedup is 1.80 by two computers parallel, and image processing time is greatly reduced. So, system scheduling methods are effective.