An Anchor-Free Network With Box Refinement and Saliency Supplement for Instance Segmentation in Remote Sensing Images

Abstract

The mainstream instance segmentation in remote sensing images takes the way of “Detect then Segment”. Anchor-free detectors get rid of the predefined anchors and can be used for proposal generation. However, in the existing methods based on anchor-free detectors, incompact proposals are directly used for segmentation, which may result in incomplete mask segmentation. In addition, the commonly used mask segmentation module is based on the trimmed features, which is not sufficient for accurate mask segmentation due to the loss of spatial details. In this letter, an anchor-free network is proposed for instance segmentation in remote sensing images. In order to obtain more compact proposals, a box refinement module, which predicts the distance offsets of a pixel to four sides of a proposal, is designed for initial proposals generated by an anchor-free detector. Based on the refined proposals, a saliency supplement module is designed to obtain accurate instance-wise masks by embedding the saliency map into the coarse masks. Compared with other methods, the proposed method conducted on two remote sensing datasets achieves optimal performance, with average precision (AP) of 0.652 and 0.678, respectively.