Edge-Based Unsupervised Evaluation of Image Segmentation

This paper focuses on a fundamental problem in computer vision: how to evaluate the quality of image segmentation. Supervised evaluation methods provide a more accurate evaluation than the unsupervised methods, but these methods cannot work without manually-segmented reference segmentations. This shortcoming limits its applications. We present an edge-based evaluation method which works without the comparison with reference segmentations. Our method evaluates the quality of segmentation by three edge-based measures: the edge fitness, the intra-region edge error and the out-of-bound error. Experimental results show that our method provides a more accurate evaluation than those method based on the statistic of pixel values, and can be used in both segmentation evaluation and region evaluation. A significant linear correlation is shown between the evaluation scores of our method and two widely used supervised methods. The proposed methods show a high performance on the automatic choice of the best fitted parameters for region growing.