Dual-stage recognition method for low-light underground coal-rock images integrating adversarial enhancement and Proxy attention

To address the low recognition accuracy of coal and rock images in underground mining environments with low illumination, high noise, and blurred boundaries, a coal and rock image recognition method integrating adversarial enhancement and Proxy Attention mechanisms is proposed. This method employs a dual U−shaped structure, forming a complete pipeline from image enhancement to semantic segmentation. The first U−shaped structure is the Coal−Enhance GAN quality enhancement module, which introduces the SRA−SA gated attention mechanism to achieve self-regularized brightness balance and spatial texture enhancement. It combines deformable convolutions and deformable RoI pooling modules to adaptively capture irregular edge structures, while a global-local discriminator with a relative discriminative strategy is used to improve the authenticity of the generated images, effectively clarifying the coal and rock images. The second U−shaped structure is the Proxy Swin−UNet semantic enhancement module, which focuses on extracting and representing high-order semantic features of the coal and rock regions from the enhanced images. The Proxy Attention module is embedded, promoting multi-scale feature interaction through proxy token bias and depthwise separable convolutions, thereby enabling precise segmentation of complex structures. Experiments on real underground coal and rock datasets show that: The enhancement module achieves an NIQE score of 3.121, improving by 6.6% and 24.4% compared to CycleGAN and RetinexNet, respectively; The segmentation module achieves an MIoU of 84.10% and an MPA of 81.91% under noise interference (noise ratio 0.05), significantly outperforming comparative models such as Swin−UNet; Ablation experiments validate the effectiveness of each module, with the complete model reducing the NIQE further to 2.874. Using the segmentation module alone under noise-free conditions, the MIoU is 85.34% and the MPA is 83.12%, which is significantly lower than the post-enhancement segmentation performance, demonstrating the crucial role of image enhancement in improving recognition accuracy. This method provides a visual perception solution for intelligent underground mining systems and holds significant engineering application value.