A hybrid approach toward model-based texture segmentation

摘要

A hybrid texture segmentation algorithm, combining statistical (maximum likelihood/maximum a posteriori) and structural (local consensus) classification techniques, is developed. This algorithm exhibits several advantages over traditional algorithms based solely on statistical models: (i) homogeneity and integrity of each region in the texture mosaic can be included implicitly by a local voting scheme, in addition to explicit modelling through Gibbs random fields, (ii) modified maximum likelihood solution of hybrid segmentation algorithm provides a better initial estimate of region boundaries which can be used in a maximum a posteriori segmentation algorithm. With this initial estimate, an iterative, semi-deterministic relaxation algorithm, called mean field annealing, is used to locate the nearly global optimum solution efficiently. The results indicate that the hybrid model provides better immunity to misclassification than the traditional single pixel window-based classification approaches.