by Tamás Szirányi and Josiane Zerubia
This article describes a new computing architecture which can process probabilistic image segmentation and labeling. This method can be implemented in real-time VLSI architectures. The architecture will allow difficult image processing methods, including probabilistic segmentation, to be implemented in single-chip analog or digital processing arrays. This image-processing array-architecture has been jointly developed at SZTAKI and at INRIA.
Markovian approaches to early vision need a huge amount of computing power. These algorithms can usually be implemented on parallel computing structures. We have shown that the Markovian labeling approach can be implemented in fully parallel cellular network architectures, using simple functions and data representations. It makes it possible to implement our model in real VLSI parallel imaging chips.
As an example, we have developed some simplified statistical image segmentation algorithms for the Cellular Neural Networks Universal Machine (CNN) (invented by T. Roska at SZTAKI and L.O. Chua at UCB Berkeley, 1993) which is a new image processing tool, implemented on a single analog chip containing several thousands of cells with analog dynamics.
The Modified Metropolis Dynamics optimization method (INRIA, 1994) can be implemented into the raw analog architecture of the CNN. We use simple arithmetic functions (addition, multiplication), equality-test between neighboring pixels and very simple nonlinear output functions (step, jigsaw). With this architecture, a real VLSI CNN chip can execute a pseudo-stochastic relaxation algorithm of about 100 iterations in about 1 msec. In the proposed solution, the segmentation is unsupervised.
We propose a general fully parallel VLSI (analog or digital) array architecture for probabilistic segmentation, which may execute several other complex preprocessing steps in the same or similar architecture, such as deblurring, texture-classification or anisotropic diffusion (SZTAKI, 1994-96). With these preprocessing steps, our fully parallel cellular system may work as a high-level image segmentation machine, using only simple cell-functions based on the close-neighborhood of a pixel.
The present results were published at CNNA96, the 4th IEEE International Workshop on Cellular Neural Networks & Applications in Sevilla, Spain, at ICPR'96, the 13th International Conference on Pattern Recognition in Vienna and in the IEEE Trans. CAS 96.
During the period 1995-96, this work was partially supported by the Balaton program of the National Committee for Technological Development of Hungary and the French Ministry of Foreign Affairs.
Tamás Szirányi - SZTAKI
Tel: +36 1 2698265
Josiane Zerubia - INRIA
Tel: +33 4 93657865