Fuzzy Evaluation of Image Segmentation Algorithms Using Neural Networks

The color and texture features are very complex in natural images, usually the segmentation algorithms cannot segments these images well and better algorithms must be chosen from among the other algorithms. In this paper we present a fuzzy novel metric to evaluate the complex images using neural network and boundary accuracy, segment-by-segment comparisons of a segmented image and a groundtruth based on fuzzy Gaussian function. The neural network after training can assess the similarity or dissimilarity of each pairs of segments and finally the segmentation algorithms accuracy have been computed by novel presented metric. Our proposed method is sensitive to over-segmentation and undersegmentation. Experimental results were obtained for a selection of images from Berkeley segmentation data set and demonstrated that it s a proper measure for comparing image segmentation algorithms.