MapReduce-based Image Processing System with Automated Parallelization

A. V. Sozykin ; M. L. Goldshtein

Vestnik Ûžno-Uralʹskogo gosudarstvennogo universiteta. Seriâ, Matematičeskoe modelirovanie i programmirovanie, no. 13 (2012), pp. 109-118

Résumé

The article describes a parallel image processing framework based on the Apache Hadoop and the MapReduce programming model. The advantage of the framework is an isolation of the details of the parallel execution from the application software developer by providing simple API to work with the image, which is loaded into memory. The main results of the work are the architecture of the Hadoop-based parallel image processing framework and the prototype implementation of this architecture. The prototype has been used to process the data from the Particle image velocimetry system (the data from the PIV challenge project have been used). Evaluation of the prototype on the four-node Hadoop cluster demonstrates near linear scalability. The results can be used in science (processing images from the physics experimental facilities, astronomical observations, and satellite pictures of a terrestrial surface), in medical research (processing images from hi-tech medical equipment), and in enterprises (analysis of data from security cameras, geographic information systems, etc.). The suggested approach provides the ability to increase the performance of image processing by using parallel computing systems, and helps to improve the work efficiency of the application developers by allowing them to concentrate on the image processing algorithms instead of the details of parallel implementation.

Détail

Keywords: image processing, MapReduce, Hadoop, distributed file system, automated parallelization.