Weakly Supervised Image Semantic Segmentation with Bounding Box Annotation

Abstract

Image semantic segmentation can understand and analyze image scenes for low-level computer vision. But it has always faced the challenges of complex feature extraction and difficult data annotation. In view of the time-consuming and expensive of pixel-level labeling, we mainly study a weakly supervised image semantic segmentation model on bounding box annotations. Firstly, the image pixel-level feature extraction is performed through a densely sampled fully convolution network based on pyramids modules, and then the GrubCut algorithm is used to process the weakly supervised data. Finally, the image features and labeled data are jointly trained to build a weakly supervised image semantic segmentation model for bounding box annotations. Experimental results show that the weakly supervised model constructed in this paper achieves better segmentation results than other weakly supervised models.