A work performed in the framework of the

Parallel Application Technology Project launched by Cray Research

Abstract

We have developped a new technique which allows interactive optimization of video compression algorithms using massively parallel computers such as the CRAY T3D. This work aims to exploit as much as possible the parallel nature of digital image processing algorithms to obtain almost real-time computing with the flexibility of a software implementation. Thanks to this low computation time, interactive tools have been developed which allow easy and fast visual evaluation of image quality. This leads to significant productivity gain when developing new video compression techniques. Our approach has been validated on advanced region-based video compression algorithms. The interactive facilities offered by the proposed technique permit the accurate optimization of the algorithm parameters in few minutes, where several days were previously needed. Depending on the complexity of the compression algorithms, 8-12 images are compressed, decompressed and visualized per second.

Introduction

One of the main difficulties for researchers when developing new digital image processing algorithms is related to the computation time required for the scientific evaluation of the algorithms, and the optimization of their parameters. As a consequence, in practice, researchers have often not enough time to optimize and to evaluate their algorithms correctly.

Due to the parallel nature of almost all image processing techniques, massively parallel computing is an attractive solution to significantly reduce the computation time. In this context, some authors have already proposed efficient parallel implementations of various video compression algorithms.

Unfortunately, even with such parallel implementations, the scientific evaluation of the results remains difficult due to the fact that there is no mathematical criteria to correctly evaluate the visual quality of an image. Such criteria are classically used to compare original images with their decompressed versions, but they are not sufficiently reliable. For this reason, in the undergoing development of the MPEG-4 video standard, the quality of the results is judged only visually. Furthermore, image processing algorithms often have several interdependant parameters which must be optimized simultaneously. This kind of problems remains unsolved even with existing parallel approaches.

In order to overcome these problems, we present in this paper an interactive system, called DirectView, for real-time video processing and real broadcasting system simulations. This system is based on efficient parallel implementations of video compression algorithms on a massively parallel computer such as the CRAY T3D with the modularity and the flexibility of a software implementation in high level languages (C, C++). Furthermore, DirectView permits the visual evaluation of the processed image sequences as they are produced by an algorithm running on the parallel computer. It also allows the modification of the parameters of an algorithm in an interactive manner, thus permitting fast and efficient evaluation and optimization. The possibility of performing fast and accurate optimization is of particular importance when considering that the algorithms used in the MPEG-2 video standards were retained mainly because they were fully optimized. It should be pointed out that real-time processing can also be obtained with dedicated hardware or with DSPs, but these kinds of implementations do not offer the flexibility given by the presented system.
its own part of the bitstream, locally compressed.

This work proves that Parallel Supercomputers can be efficiently used for the fast development of new video processing algorithms by exploiting their intrinsicly parallel nature. In this context, the technology developed with DirectView offers real-time interactive optimization facilities with the advantages of flexible and modular C software implementations. Better results are obtained both in terms of compression ratio and in terms of visual quality of the decompressed images, compared to classical approaches based on mathematical optimization criteria. Furthermore, DirectView can be used to develop and to optimize a wide range of video processing algorithms, thus permitting quick prototyping and significant productivity gain.

This snapshot shows what sees the operator when optimizing its compression algorithm:

Several critical parameters of algorithm with runs on Cray T3D,

The Real-time decompressed image (10 frame/s in CIF format),

The bitrate variation which allows to get a given trade-off between bitrate and quality.

Feel free to send an email if you are interested by this kind of application.