A Survey of Weakly-supervised Image Semantic Segmentation Based on Image-level Labels

Abstract

According to the different ways of image-level label location inference, the weakly-supervised image semantic segmentation methods with image-level labels were divided into superpixel-based methods and classification-network-prior based methods. Then, various methods were discussed and summarized in detail from the principles, advantages and disadvantages, key links, main technologies, features, superpixel/candidate region segmentation, seed region generation, network structure and dataset, etc. Second, the commonly used datasets and evaluation indexes were described for weakly-supervised image semantic segmentation based on image-level labels, and the characteristics of each data set were introduced. Finally, the performance of weakly-supervised image semantic segmentation methods was compared on the basis of image-level labels on MSRC, PASCAL VOC 2012, MS COCO, and Sift Flow datasets. Moreover, the research directions of weakly-supervised image semantic segmentation were prospected from the large-scale multimedia sharing website, specific application scenarios, and strategies of image-level label location inference.

Keywords

• semantic segmentation
• weakly-supervised learning
• image-level labels
• superpixels
• deep convolutional neural network