Research on recognition of abnormal areas in infrared thermal images of coal and rock failure based on deep learning

To recognize failure areas in coal and rock using infrared thermal imaging, an improved U-Net network is proposed to segment abnormal areas in the images. This method increases the segmentation accuracy of failure areas in coal and rock. Meanwhile, to improve the segmentation effectiveness of infrared thermal images, a dense residual denoising algorithm is introduced based on autocorrelation networks to preprocess the images before segmentation. The results show that the denoising algorithm maintains details and structural features in infrared thermal images while reducing edge information loss and artifacts. It also significantly improves the clarity of infrared thermal images, and effectively increases the segmentation accuracy of the segmentation model for infrared thermal images. Its accuracy, F1 score, Dice coefficient, and MIoU value are improved by 0.86 %, 0.37 %, 1.07 %, and 1.55 %, respectively, and the training time is shortened by 12.28 %. Compared with other deep learning models, the improved U-Net network has a higher performance, with its accuracy reaching 94.36 %, F1 score reaching 94.11 %, Dice coefficient reaching 91.82 %, and MIoU value reaching 86.93 %. Combining the two algorithms supports the improvement of accurate identification of abnormal areas in coal and rock failure using infrared thermal images. This research paves the way for intelligent monitoring and early-warning systems for coal and rock dynamic disasters.