Coal-enhance GAN: Low-light enhancement model for intelligent coal-rock recognition

In underground coal mines, the extremely low illumination, high dust concentration, and mechanical vibration result in coal-rock images suffering from triple degradation: grayscale collapse, edge diffusion, and structural fragmentation. These issues severely hinder the accuracy of visual recognition in intelligent mining applications. To address the challenge that existing methods struggle to enhance brightness while preserving fine details, this paper proposes a novel deformable attention-guided generative adversarial network, termed Coal-Enhance GAN. The main innovations are as follows: (1) Illumination-Texture Decoupled Enhancement Mechanism: A hybrid SRA-SA gating module is designed, integrating Self-Regularized Attention (SRA) and Spatial Attention (SA). This module dynamically balances enhancement in dark regions and suppression in bright areas through luminance-adaptive weighting, thereby mitigating halo artifacts common in traditional Retinex-based methods. (2) Deformation-Adaptive Feature Extraction: Deformable convolution and deformable RoI pooling are introduced to capture irregular edges caused by dust and vibration through learned offset fields, enhancing robustness against geometric distortions. (3) Multi-Scale Discriminative Constraints: A global-local discriminator architecture is developed, along with a relative adversarial loss, to jointly optimize global illumination uniformity and local structural realism (e.g., coal-rock fractures), effectively addressing the overexposure artifacts often seen in CycleGAN. Experimental results on 500 underground coal-rock test images show that the proposed method achieves a NIQE score of 3.124, significantly outperforming CycleGAN (3.338) and RetinexNet (4.126). Ablation studies confirm the synergistic effectiveness of the proposed modules: the complete model reduces NIQE by 30.9. Edge sharpness is quantified using the Average Gradient (AG), a widely accepted metric for evaluating image clarity by measuring intensity variations across pixel neighborhoods. AG is computed as: AG = 1 ( M − 1 ) ( N − 1 ) ∑ i = 1 M − 1 ∑ j = 1 N − 1 G x 2 ( i , j ) + G y 2 ( i , j ) where G x and G y are horizontal and vertical gradients calculated using Sobel operators, and M , N represent image height and width. This formulation aligns with standard practices in image sharpness assessment, where higher AG values indicate clearer edge details. In experiments, the baseline model (without deformable components) yielded an AG value of 0.321, while integrating deformable components increased AG to 0.468. This corresponds to a relative improvement of 0.468 − 0.321 0.321 × 100 % ≈ 46 % . The observed AG values are consistent with the low-illumination characteristics of underground coal-rock images, where inherent noise and uneven lighting typically result in lower baseline sharpness metrics compared to well-lit natural scenes. This model offers a high-fidelity visual preprocessing solution for intelligent coal-rock recognition and sorting, advancing the practical deployment of smart mining systems.