| In this dissertation, we study the application of a learning and adaptive systems to image restoration problem. This thesis is divided into three parts. In the first part, we focus on a specialized learning system—the projection pursuit learning network (PPLN). We study its capabilities and limitations in the context of noisy and blurred images. A comparison with the commonly used Wiener filter is included to put the learning system's performance in perspective.; In the second part, we digress to get a better understanding of the role of smoothing functions in image restoration problem. Here, we propose a family of exponential functions to smooth noisy images. We observe that optimal results are obtained when the value of the exponent lies within a certain range. We establish that the range found is implicitly present in the structure of all natural images. Furthermore, we discover that this range has been found to be useful in applications other than image smoothing.; In the last part, we consider modular learning network for image restoration problem. We propose a hybrid learning method to combine unsupervised and supervised learning algorithms. We study the complex task of input clustering, a range of gating functions and a novel way to extend the role of input to achieve superior restored images. A new quantitative measurement technique to rank the restored images, that agrees with visual inspection of images, is proposed here. |
