Fusion attention-based model for low-light image enhancement in mines

To address the problems of dark regions, insufficient detail recovery, and simultaneous noise amplification that easily occur during underground mine image enhancement, this study proposed a fusion attention-based low-light image enhancement model for mines. An encoder-decoder network was adopted as the backbone network, and Residual Dense Block (RDB) was introduced at each stage to alleviate the attenuation of weak texture information in dark mine regions during deep feature transmission and the difficulty in preserving details. A Channel-Spatial Attention (CSA) module was designed in the decoding stage to enhance the responses of key regions such as equipment edges and personnel contours and to suppress interference from background noise and dust scattering. A composite loss function consisting of mean squared error loss and perceptual loss was constructed to address the difficulty of balancing structural preservation and visual naturalness during brightness recovery, thereby improving structural fidelity and visual quality while ensuring pixel consistency. Comparative experiments were conducted on the self-built CMUHL mine dataset. The results showed that the proposed model achieved a Peak Signal-to-Noise Ratio (PSNR) of 30.544 dB and a Structural Similarity Index (SSIM) of 0.914, outperforming FLOL, AnlightenDiff, Retinexformer, and other models. The model had 4.37×10⁶ parameters and a computational cost of 56.19 GFLOPs, indicating that it achieved superior enhancement performance while balancing model complexity and deployment feasibility. Real-scene validation results showed that the proposed model achieved the lowest Natural Image Quality Evaluator (NIQE) value, indicating better statistical naturalness of the enhanced results. The Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) values remained relatively low overall and were consistent with the subjective visual effects, indicating that the model had good stability and applicability in complex real underground mine lighting environments.