Fuzzy Evaluations of Image Segmentations

Evaluation measures for images segmentation are suggested. The methods compare the results of automatic segmentation with ground truth. The presented methods for assessing the similarity of the segments are based on three different approaches: the number of pixels in common, the similarity of the contours, and the location of centroids. The fuzzy approach consists of considering the significance of segment differences in relation to the size of the segments. The final measures for the whole images are based on recall and precision, widely used in information retrieval tasks. The approaches presented in this paper apply the fuzzy set theory instead of classical evaluation methods.