Instance segmentation of coarse aggregate images based on improved SOLOv2

Abstract This article proposes an improved SOLOv2 instance segmentation algorithm, which solves the problems of missed detection, false detection, and poor segmentation accuracy of coarse aggregates under complex working conditions such as dryness, moisture, dispersion, adhesion, and stacking by introducing a cascaded channel spatial attention module before the mask branch. The experimental results show that the improved SOLOv2 model is more accurate in boundary segmentation of coarse aggregate instances under complex working conditions. Compared with the original SOLOv2, YOLACT, and Mask R-CNN models, the mAP has increased by 1.2%, 2.6%, and 2.0%, respectively.