A neural learning approach for adaptive image restoration using a fuzzy model-based network architecture

In this paper, we address the problem of adaptive regularization in image r estoration by adopting a neural-network learning approach. Instead of expli citly specifying the local regularization parameter values, they are regard ed as network weights which are then modified through the supply of appropr iate training examples. The desired response of the network is in the form of a gray level value estimate of the current pixel using weighted order st atistic (WOS) filter. However, instead of replacing the previous value with this estimate, this is used to modify the network weights, or equivalently , the regularization parameters such that the restored gray level value pro duced by the network is closer to this desired response. In this way, the s ingle WOS estimation scheme can allow appropriate parameter values to emerg e under different noise conditions, rather than requiring their explicit se lection in each occasion. In addition, we also consider the separate regula rization of edges and textures due to their different noise masking capabil ities. This in turn requires discriminating between these two feature types . Due to the inability of conventional local variance measures to distingui sh these two high variance features, we propose the new edge-texture charac terization (ETC) measure which performs this discrimination based on a scal ar value only. This is then incorporated into a fuzzified form of the previ ous neural network which determines the degree of membership of each high v ariance pixel in two fuzzy sets, the EDGE and TEXTURE fuzzy sets, from the local ETC value, and then evaluates the appropriate regularization paramete r by appropriately combining these two membership function values.