Algorithms on generating neighborhood systems in a constrained optimization framework for image segmentation

In the past decades, image segmentation has been known to be an important process for scene analysis and image understanding in the area of computer vision and pattern recognition. In this dissertation, we consider an unsupervised constrained optimization framework for image segmentation, based on Gibbs random field model. The input image is considered as a collection of a number of cells of the same resolution. It is partitioned into homogeneous regions with assigned generic labels, based on similarity between regions in the feature space. It does not require any a priori information for specific texture samples. When multiple distinct features regarding the texture information of the image are employed, they may be combined via Dempster-Shafer theory of evidence. Certain forbidden label configurations are determined based on domain knowledge. A stochastic relaxation algorithm with simulated annealing is used for building an optimal label map such that the energy function of the entire image, accounting for the penalty of constraints, is minimized.; Among various components of the proposed approach, we particularly focus on the generation of neighborhood systems. The neighborhood system is substantially important not only in the sense that it characterizes the spatial and statistical dependence of the image data, but also the energy function of the entire image. Frequently employed are nearest neighborhood systems of low orders, which may yield an incorrect separation of distinct textures in our segmentation scheme. Hence, as an improvement, a number of neighbors chosen randomly at varying spatial ranges are required. For serving this purpose, three algorithms, Vertex Scan, Random Trail, and Structural Recursion, are presented, based on the analogy with graphs and fundamental concepts of graph theory. The performance of each algorithm depends on the number of sites and the number of neighbors associated with each site. Comparative analysis among these algorithms is given.; Experimental results on several synthetic digital fractal images have demonstrated that a neighborhood system containing several nearest as well as random neighbors performs very well. The performance of the described segmentation scheme has also been evaluated for the detection of tumor mass on 35 digital mammograms that contain various types of abnormalities. The results are quite encouraging and validate the effectiveness of our approach. The detection rate for mammograms with a malignant mass is 84.21%, while the average number of false alarms for each mammogram with detected abnormality is less than one.