SRD-NET for Ground Crack Detection in Coal Mines Using UAV Images

Abstract

The large-scale coal exploitation causes numerous surface cracks in mining zones. These cracks endanger area safety, damage the ecological environment, and threaten local people’s lives. Traditional ground survey methods for crack detection are inefficient, costly, and limited, failing to meet monitoring demands. To address this, this study uses drone images and deep learning to identify ground cracks. An enhanced model, SRD-NET, based on U-NET, is proposed. It incorporates SE, DSC, and residual connections to improve crack feature recognition and generalization. Experimental results on a dataset of 400 $512\times 512$ -pixel images collected from Huipodi Coal Mine, where 210 were for training, 60 for validation, and 30 for testing, demonstrate the model’s outstanding performance. Compared with U-NET, SRD-NET’s mPrecision is 5.6% higher, mRecall is 10.56% higher, mF1 is 7.16% higher, and mIoU is 7.14% higher. Against DSC-NET, SRD-NET’s mPrecision is 6.93% higher, mRecall is 11.46% higher, mF1 is 8.31% higher, and mIoU is 8.41% higher. When compared with residual network (Res-Net), SRD-NET’s mPrecision is 3.71% higher, mRecall is 9.00% higher, mF1 is 5.18% higher, and mIoU is 4.99% higher. Although SRD-NET’s mPrecision, mRecall, mF1, and mIoU are 0.38%, 0.15%, 1.37%, and 0.45% lower than SR-NET, respectively, SRD-NET’s FPS is 44 and 6 frames/s higher than SR-NET. Overall, SRD-NET improves the segmentation accuracy and has a relatively high processing speed, effectively demonstrating its efficacy in ground crack identification tasks.