A fuzzy approach to evaluate image segmentation based on image complexity

Abstract

Image segmentation evaluation may be carried out by comparing segmentation algorithm results with human ground truths. Its correct evaluation in benchmarks is important to promote the development of new and better segmentation algorithms. It may be partially considered a subjective task because each image may have multiple correct solutions. The evaluation is commonly carried out through crisp metrics, and they fail to generalize the subjectivity of the human criteria present in the human ground truths. Therefore, the interpretation and meaning of the metric values may be ambiguous. Thus, this paper presents a new fuzzy evaluation approach that considers the subjectivity of the human criteria by considering image complexity. This approach leads to an adequate evaluation of this subjective task. The proposed approach demonstrates its advantage when is evaluated using the Peng and SIHD meta-metrics achieving a performance of 0.795 and 0.971, respectively, outperforming the PRI, VI, GCE, and BDE metrics.