Segmentation Quality Evaluation Algorithm Based on Classification

Abstract

Image segmentation plays a vital role in computer vision. Image segmentation quality evaluation is an essential task in image segmentation and an important step to quantify the performance of the segmentation algorithm. Most of the existing evaluation methods need to use ground truth to evaluate the segmentation quality. However, the annotation of ground truth varies from person to person and takes a long time. In this paper, we proposed a new segmentation quality evaluation network. In the evaluation of segmentation quality, only the segmentation results to be evaluated and the original image are required without specially annotated ground truth. At the same time, we also propose a new space to squeeze module (STS) for segmentation quality evaluation. STS module autonomously learns the edge features of the segmentation object and increases the weight of edge features. Experiments on the dataset constructed in this paper show that the performance of the proposed network is better than other network structures such as ECA-Net, CBAM, SE-Net, and the evaluation accuracy is higher than the existing supervised and unsupervised segmentation quality evaluation methods.