Comparative Analysis of Efficient Platforms: Scalable Algorithms and Parallel Paradigms for Large Scale Image Processing

1. Introduction

This research is part of an ongoing project that aims to offer a service of dematerialization of documents to very small enterprises. Dematerialization is being done in order to archive documents that exist in a paper form, and efficiently retrieve them on demand from anywhere. When this service will be operational, scanners located in very small firms will be able to digitize paper documents before sending these so obtained numeric images to a server. Then, this server will process the received images in order, on one hand, to improve the quality of and, on the other hand, to reduce the volume of the digitized images. This process has been described in Figure 1.

Figure 1.

Conversion and archival of documents

Finally, the processed images will be communicated to an archiving center which will allow the owner of an archived document, to consult its digital image. For this purpose data and task parallelism is required for High Performance Computing and the processing of every image may contain tasks which could be run in parallel on the various cores of each computational node. Image processing is widely applied in many applications such as medical image processing, non-photorealistic rendering, remote sensing, optical sorting and many more. The processing of the flow of images is difficult to characterize. Some applications require to process images that are of very large size and require that all the processing can be done in a very short interval of time. The throughput of arrival of the images may undergo great and unpredictable variations according to the time and the duration of the processing of an image can also undergo great and unpredictable variations according to the image type and size. The overall process can be divided into three portions scanning and transmitting, processing and archival. The focus of this study is towards the processing module which receives images and processes them. Our aim is to study the problems that may arise in this area and technologies that address them, and propose and test a solution. Following are some areas of concern: Processing of images that are contained in a stream. The number of images to process at any given time may vary to a great extent The processing time is dynamic and depends on each individual image Because of the varying amount of load on the processing module, a scalable platform is needed to be identified and technologies that provide parallelism have to be studied so that the system efficiently processes high number of images in a certain amount of time. This chapter describes various aspects of different technologies related to image processing on hybrid platform. The objective of this research is to analyze various technical solutions in parallel image processing, where different techniques are compared to each other, and the best suited techniques are suggested for parallel image processing on large scalable infrastructure. Section 2 analyzes different platforms and argues for the suitability of one of them on the basis of different measures and benefits. Section 3 presents different programming languages and libraries for parallel image processing (task parallelism), while section 4 discusses few techniques and algorithms for data parallelism. Section 5 gives the detail of the experiments conducted and test the environment. Section 6 concludes and suggests the future directions.