Restoration and segmentation of digital images by adaptive filtering

Segmentation of degraded images has always been a difficult problem to solve. In coherent image acquisition systems, occurrence of speckle noise is a common phenomenon that is hard to remove without further degrading the image. In this dissertation, a new method for image segmentation based on the Adaptive Fuzzy Leader Clustering Algorithm (AFLC) is introduced. AFLC is a hybrid neuro-fuzzy model developed by integrating a Learning Vector Quantization (LVQ) network with fuzzy memberships. This integration provides a powerful yet fast method for recognizing embedded data structure and it has shown superior misclassification rates over similar segmentation approaches, Neuro-fuzzy clustering algorithms can achieve efficient object extraction from noisy images since noise pixels can be identified during the clustering process and separated from the rest of the image. When dealing with corrupted images, the first step prior to segmentation is always the enhancement of features in the image by filtering in either the spatial or frequency domain. In this dissertation, a new non-linear adaptive filter based on AFLC is developed. This new adaptive filtering method has been specifically tailored to reduce the degradation introduced by speckle noise in coherent imagery like synthetic aperture radar (SAR) or ultrasound imaging. The results achieved by this process have been compared with the results from the traditional median filter, the Kuan filter, and the connectivitypreserving morphological filter demonstrating the superior performance of AFLC in removing speckle noise. We have also compared AFLC to other classification algorithms, such as those derived from the statistical decision theory, and to many well-known fuzzy, neural, and neuro-fuzzy unsupervised algorithms. The concept of local cluster validity introduced in AFLC is addressed and comparison results are presented for well-known global validity indices. Finally, the convergence criteria of the AFLC algorithm have been analyzed under the framework of stochastic approximation and results that ensure the stability of its performance are presented.