Abnormality Detection by Generating Random Fields Based on Markov Random Field Theory

Image segmentation plays an important role in abnormality detection. In difficult image segmentation problems, multidimensional feature vectors from filter banks provide effective classification within homogeneous regions. However, such band limited feature vectors often exhibit transitory errors at the boundaries between two regions. At boundaries, the feature vector may make a transition through a region of feature space that is incorrectly assigned to a third class. To remove such errors, a new method is proposed based on binary random variables to eliminate boundary errors. The proposed method for eliminating the narrow misclassified regions proceeds in two steps, In the first step, pixels in the classified image whose neighborhood consists entirely of one class are left unchanged; otherwise, the pixel value is set to zero to indicate that the pixel is no longer assigned to any class. In the second step, the classified regions are propagated back into the unassigned regions based on the most common class within neighborhood system. The significant improvement is obtained compared to the traditional