Coal-rock interface image recognition based on MRU-Net++ for extremely thin coal seam fully-mechanized mining face

Abstract

Coal rock recognition is one of the core technologies to realize intelligent mining in the integrated mining face of extremely thin coal seam. Aiming at the special situation that the coal-rock boundary is naturally exposed during the mining of extremely thin coal seams, an image recognition method based on MRU-Net++ network was proposed for coal-rock image recognition of extremely thin coal seams. The network is based on U-Net++, and the structure of U-Net++ was optimized by the method of “pruning”, which reduces the complexity of U-Net++ network while minimizing the loss of its performance in order to improve the computing speed. MobileNetV2 lightweight network was used to construct a core backbone network based on MobileNetV2, replacing the original network architecture of U-Net++, which significantly reduces the number of parameters of the model and improves the efficiency of the model segmentation. At the same time, the ResNeSt module, which contains the channel attention mechanism, was introduced to enhance the ability of extracting the detailed features of the edges of the coal and rock images, and to increase the segmentation accuracy. The explosion-proof camera on the hydraulic support was used to collect the coal rock images of the comprehensive mining face in the extremely thin coal seam, and the high-definition coal rock images with coal rock distribution information were acquired and preprocessed to create a coal rock image dataset of the comprehensive mining face of the extremely thin coal seam containing 2 536 samples. The ablation test was set up to verify the effect of the improved part on the network performance. The model was compared with the classical FCN, U-Net, and U-Net++ network models and the network models were trained using adaptive learning algorithms. Key indexes such as Pixel Accuracy (PA), Intersection over Union (IOU) and test time were selected to evaluate the model segmentation effect. The results show that the Mean Pixel Ascuracy (PAM) and Mean Intersection over Union (IOUM) of the MRU-Net++ network model are 97.15% and 94.16%, respectively, the memory occupied by the model is 25.71 M, and the average test time of each image is 28.61 ms, which fully proves the feasibility and effectiveness of the method for the coal rock recognition task under the special environment of extremely thin coal seam.

Keywords

• extremely thin coal seam
• coal rock recognition
• semantic segmentation
• u-net++ network model
• deep learning