CNN architectures for constrained diffusion based locally adaptive image processing

In this paper, a cellular neural network (CNN) based locally adaptive scheme is presented for image segmentation and edge detection. It is shown that combining a constrained (linear or non-linear) diffusion approach with adaptive morphology leads to a robust segmentation algorithm for an important class of image models. These images are comprised of simple geometrical objects, each having a homogeneous grey-scale level and they might be overlapping. The background illumination is inhomogeneous, the objects are corrupted by additive Gaussian noise and possibly blurred by low-pass-filtering-type effects. Typically, this class has a multimodal (in most cases bimodal) image histogram and no special (easily exploitable) characteristics in the frequency domain. The synthesized analogic (analog and logic) CNN algorithm combines a diffusion-type filtering with a locally adaptive strategy based on estimating the first-order (mean) and second-order (variance) statistics. Both PDE- and non-PDE-related diffusion schemes are examined and compared in the CNN framework. It is shown that the proposed algorithm with various diffusion-type filters offers a more robust solution than some globally optimal thresholding schemes. All algorithmic steps are realized using nearest-neighbour CNN templates. The VLSI implementation complexity and some robustness issues are carefully analysed and discussed in detail. A number of tests have been completed on original and artificial grey-scale images. Copyright (C) 2002 John Wiley Sons, Ltd.