A high-confidence instance boundary regression approach and its application in coal-gangue separation

Coal-gangue sorting robots have been established as a ground-breaking technology in mineral processing, particularly for mitigating labor intensity and increasing production efficiency. Existing coal-gangue sorting robots tend to employ object detection algorithms as the framework, but the crude bounding boxes provided by this approach is insufficient to support the robot in all sorting scenarios. The instance boundary enables the robot to dexterously grab the target amid numerous interfering objects, which is essential for enhancing the capabilities of the coal-gangue sorting robots. Currently, there is an increasing focus on the performance of instance segmentation algorithms at boundaries, but their mindset is still to extract the outer edge based on the segmentation instance mask, which is unable to determine whether the output boundary is reliable. In this study, we proposed a framework for generating high-confidence instance boundary and named it BGF. The overall idea is to crop each instance into patch based on output of object detection model; the boundary regression network then directly generated the boundaries of the salient target in the patch; all results were finally reassembled to get a global instance boundary with detailed semantic information. In parallel, we introduced an attention mechanism and auxiliary branch to enhance the model's ability on extracting boundary features. The rationality of various designing in the BGF has also been demonstrated through a serious of experiments on varies dataset. Compared with conventional segmentation networks, BGF has more credible capability in boundary extraction and generalization.