A multi-module fusion network for coal-rock interface recognition

Accurately identifying the coal-rock interface is an effective path to enhance coal mining efficiency, guarantee coal quality, and reduce safety accidents. Due to the small differences in color and texture of coal rock, especially the complex texture, blurred boundary and irregular shape at the coal-rock junction, and the existence of interference factors such as low light and dust in the mine, the under-segmentation and mis-segmentation problem appear in the identification of coal-rock interface. In addition, the large number of parameters in the recognition model result in a low recognition efficiency. Therefore, this paper proposes a coal rock image segmentation network named CBAM-SP-CARAFE-DeepLabV3+ based on the deep learning technology. Firstly, the network selects MobileNetV2 as the backbone and embeds the Convolutional Block Attention Module (CBAM) into the Inverted Residual Block to achieve multi-dimensional extraction of coal and rock feature information; Secondly, the global average pooling in the Atrous Spatial Pyramid Pooling (ASPP) structure is replaced with Strip Pooling(SP), and SP is concatenated into the decoder to widen the receptive field and ensure the continuity of coal rock boundary information; Finally, the Content-Aware ReAssembly of FEatures(CARAFE) is used to upsample the feature map to preserve coal and rock detail information. Experiments are conducted on the proposed network based on a self-made coal and rock dataset. The results suggest that the proposed network with fewer parameters has achieved a Pixel Accuracy (PA) of 95.45 % and a mean Intersection over Union (mIoU) of 81.32 %, which can achieve a better performance than some classical segmentation networks and some recent coal rock segmentation models.