EFASNet: An Adaptive Multi-Scale Segmentation Network for Surface Crack Extraction in Coal-Mining Subsidence Areas Using UAV Imagery

A large number of surface cracks are prone to be induced by coal mining activities in mining subsidence areas. The rapid and accurate extraction of surface cracks is of great significance for surface subsidence monitoring and safe coal mine production. To address the problems of limited surface crack samples, high labeling costs, susceptibility to background noise interference, and low extraction accuracy caused by the diverse scales of surface cracks, an adaptive surface crack extraction model considering background noise interference and multi-scale characteristics was developed to achieve the intelligent extraction of multi-scale surface cracks. To alleviate the problem of insufficient surface crack samples in mining subsidence areas, a surface-crack sample augmentation method based on an improved Conditional Generative Adversarial Network was proposed. To improve the extraction performance for multi-scale surface cracks, Atrous Spatial Pyramid Pooling, a Feature Distillation Block, and an Efficient Multi-scale Attention mechanism were introduced into the encoder. UAV imagery acquired from the Lijiahao and Jingbao coal mines in Inner Mongolia was utilized, and the effectiveness of the proposed method was verified through comparative model analysis experiments. The results indicate that texture interference, vegetation shadows, and other background noises can be effectively suppressed by the proposed method under complex background conditions. Meanwhile, the continuity and integrity of surface cracks at different scales can be significantly improved. An Intersection over Union of 85.53% was achieved, which outperformed representative models such as DeepCrack and U-Net++.