Analysis of Evaluation Metrics for Image Segmentation

Objective and quantitative evaluation for segmentation performance is important for the development of image segmentation algorithms. Several objective evaluation metrics have been proposed in the literature. This paper presents an analysis of the existing pixel-based and object-based evaluation metrics. We define and describe the possible error types of image segmentation. We investigate the properties of the error metrics by mathematical proof and experimental justification. The results indicate that the object-based metrics have several shortages although they are more suitable than the pixel-based metric for object-level evaluation.