The aim of blind image deconvolution is to reconstruct the original scene from a degraded observation without using information about the true image and the point spread function. The restoration process is critical in applications, where the true image or its statistical characteristics are unknown. Mapping the pixels of the original image before its transmission, the mapped image can be interpreted as a pulse amplitude modulation (PAM) signal, used in communications systems. With this

more »

... tation, classic equalization techniques of communication channels can be used to image restoration. Furthermore, the pixels of a true image constitute a nonstationary signal, which justies the use of adaptive lters. In this dissertation, adaptive techniques used for equalization of communication channels are applied to image restoration. Firstly, we propose a new update path through the blurred image that consists in a combination of horizontal and vertical alternate paths. This update path minimizes the problem of abrupt changes in the adaptation of the lter and provides better conditions to the image recovery. Using the least mean squares (LMS) algorithm, we obtain an equivalence between a point spread function and a time-variant communication channel. This equivalence was used to compare some point spread functions in relation to the distortion caused in images. Secondly, reshaping the input matrix into a column vector, we extend the regional-based multimodulus algorithm (RMMA) to blind image deconvolution. This algorithm is used to update the coecients of the linear transversal equalizer and also of the decision feedback equalizer. RMMA treats nonconstant modulus signals as constant modulus ones, which provides a behavior closer to that of a supervised algorithm. Thus, RMMA can converge in the mean to the Wiener solution and, therefore, presents a better performance when compared to the conventional constant modulus algorithm (CMA), used in blind equalization of communication channels. This behavior was also observed in image restoration, through the simulations presented in this dissertation. This study pushes back the frontiers of image processing, since dierent techniques used in equalization can be extended to image restoration. One of the new possibilities is the color image restoration using the spatial diversity.