Partner und Internationale Organisationen
(Englisch)
|
IMSE-CNM-CSIC, EPFL, K.U. LEUVEN, MTA. SZTAKI, ST
|
Abstract
(Englisch)
|
DICTAM is the acronym for Dynamic Image Coding using Tera-Speed Analogic Visual Microprocessors. Its primary objective is to develop a new class of image-processing systems to make hand held videocompression possible. These systems are based on the hybridization of conventional, accurate, serial DSP, and unconventional moderate-accuracy, parallel Analog Signal Programmable Processing Arrays (ASPPAs). DICTAM aims to realize a smart synergy of DSPs and ASPPAs based on an adaptive optimum distribution of the image-processing tasks among these two types of processors. In this manner the best of the two processing alternatives can be exploited and their limitations avoided. The objectives of DICTAM include the realization of the new processing architectures themselves, the conception and design of its unconventional parts in the form of mixed-signal VLSI chips, and the development of optimum signal processing algorithms, tailored to the processing architecture, for realtime dynamic image coding/decoding, video compression/decompression, and autenticity and integrity verification of the transmitted images. These global objectives require a multidisciplinary approach, involving research activities in microelectronics, signal processing and computing architectures.
In this project, LTS has been the key reference for the signal processing tasks. Its major role was to develop algorithms for dynamic image coding/decoding and video compression/decompression. In the first phase of the project the main target has been to provide an unsupervised Spatio-Temporal Image Segmentation procedure to be run with the new hardware architecture. Unsupervised image segmentation is the first step of the new video coding standard, MPEG-4. It requires a huge amount of computation power that limits its use in real time applications. In narrow collaboration with the hungarian partner, LTS has developped an algorithmic solution, which has been the subject of a conference publication [1]. This algorithm has been successively tested on the first chip prototype.
Another computation heavy task in computer vision is motion estimation. Motion is an important information in video coding, since it allows the coding procedure to exploit the temporal redundance present in the sequence. However it is wether a low accuracy or high time consuming step. LTS has investigated and proposed new algorithms compatible with the characteristics of the CNN processor. These have been the subject of an international journal publication [2].
[1] K. Laszlo, F. Ziliani, T. Roska, M. Kunt, 'Early segmentation in video compression using CNN processors'. Proceedings of 1998 fifth international Workshop on Cellular Neural Networks and theirs applications (CNNA'98), pp. 175-180, London, UK, April 14-17, 1998.
[2] T. Sziranyi, K. Laszlo, L. Czuni, F. Ziliani, 'Object oriented motion-segmentation for video-compression in the CNN-UM ' In Journal of VLSI Signal Processing Systems for Signal, Image, and Video Technology, Vol. 23, (2-3) pp. 479-496 Nov. 1999.
|
