RA-UNet++ based image segmentation of adherent ores

Abstract

Ore particle size information is an important indicator for evaluating the operating status and production efficiency of crushers. However, in the actual industrial environment, adhesion phenomena often occur during the acquisition and transportation of ores, resulting in overlapping edges and blurred contours of ores in the images. Traditional image segmentation methods are difficult to achieve high-precision recognition and segmentation. To this end, this paper proposes an RA-UNet++ adhered ore image segmentation method based on the improved UNet++ structure to improve the segmentation performance in complex scenes. Based on the UNet++ coender-decoding architecture, the residual module and the self-attention mechanism are integrated to enhance the model’s ability to extract and express the edge details of ores. Meanwhile, multi-scale atrous convolution is introduced at the end of the encoder to construct the Atrous Spatial Pyramid Pooling (ASPP) structure, expand the receptive field, enhance the multi-scale perception ability for ores of different particle sizes, and thereby improve the overall segmentation effect. The experimental results show that RA-UNet++ performs excellently in the task of image segmentation of adhered ores, significantly improving the clarity and integrity of edge segmentation. Compared with the Otsu method and the standard UNet model, this method has advantages in terms of robustness, boundary preservation, and segmentation accuracy. The pixel-level and target-level segmentation accuracy rates of adhered ore images both exceed 93%, showing good potential for industrial applications.