A supervised evaluation method based on region shape descriptor for image segmentation algorithm

In this paper we present a new supervised evaluation method for measuring the accuracy of image segmentation algorithms. This method calculates the extent of similarity between segmented images against ground truth. Feature vectors are computed based on the shape descriptor for each region in segmented image and then compared with the feature vectors of ground truth image. The proposed method can be used for any type of grayscale and color images with any number of regions. It also limits under-segmentation and over-segmentation problems. We compare the efficiency of the proposed method with extended version of four different supervised evaluation measures such as, global consistency error (GCE), local consistency error (LCE), object-level consistency error (OCE using Dice's coefficient) and the Jaccard index. Analysis of the experimental results on a large variety of test images from the Berkeley segmentation dataset demonstrates the efficiency of the proposed method.