A Diverse Environment Coal Gangue Image Segmentation Model Combining Improved U-Net and Semi-supervised Automatic Annotation

The problem of uneven illumination in the coal gangue image and the existence of fine coal gangue reduces the accuracy of the coal gangue location, which also makes it difficult to the annotation of the coal gangue images. In this paper, we propose an improved U-Net gangue segmentation algorithm to achieve gangue sorting and automatic image annotation under diversified environments. Specifically, to solve the problem that some fine coal and gangue fragments are difficult to be recognized, the InceptionV1 module is introduced instead of some of these convolutional blocks. Secondly, to solve the visual blur caused by light changes in industrial sites, the CPAM attention module is added to the fusion part of U-Net. Finally, to alleviate the difficulty of coal gangue image annotation, a semi-supervised human-in-the-loop automatic annotation framework is constructed, and a statistical method based on classification probability is proposed to automatically screen the questionable labels. The experimental results show that the proposed improved segmentation model exhibits good performance in the coal gangue sorting task with MIoU values of 80.23% and MPA values of 86.16%, which have higher accuracy compared to the original model. In addition, the statistical method proposed in this study can effectively screen out the incorrect and correct labeled data, thus enabling automatic annotation and screening.