Coal-Rock Image Recognition Method for Complex and Harsh Environment in Coal Mine Using Deep Learning Models

The unfavorable factors of underground coal such as dark light, uneven illumination, band shadowing greatly make it difficult to recognize the coal rock at the mining workface accurately. To solve this problem, this paper proposes the fuse attention mechanism's coal rock full-scale network (FAM-CRFSN) model. The deep extraction of coal rock semantic features is achieved by a multi-channel residual attention mechanism and a full-scale connection structure. Meanwhile, the balance between "deep" stacking and error back propagation is achieved by structures such as dilated convolution and Res2Block. Besides, a multi-dimensional loss function consisting of the cross-entropy loss, intersection over union, and multiscale structure similarity loss with pixel-level, area-level, and image-level expressions is established. Finally, the performance of the FAM-CRFSN network is tested with RGB coal rock images collected from an underground coal mining workface and superimposed with different proportions of gaussian noise and salt & pepper noise. The experimental results show that the FAM-CRFSN model can segment the coal rock regions accurately; at a noise intensity of 0.09, it achieves an MIOU of 85.77% and an MPA of 92.12%. Also, it achieves better accuracy and generalization performance than the mainstream semantic segmentation models. This study provides an important theoretical basis for promotes the unmanned and intelligent mining workface.