Abstract: In the complex environment of underground coal mines, images often suffer from insufficient illumination, non-uniform lighting, dust scattering blurring, and noise interference, which seriously affect the stability of monitoring and early warning as well as intelligent recognition systems. To address these issues, an image enhancement algorithm for low-illumination mine images based on an improved fusion attention convolutional network is proposed. This method employs an encoder-decoder structure as the backbone network and introduces residual dense blocks (RDB) at each stage to enhance feature reusability and alleviate gradient degradation through multi-level feature concatenation and local residual fusion. In the decoding stage, a channel-space fusion attention module (CSA) is designed, combining efficient channel attention (ECA) and depthwise separable spatial attention (DSA), to achieve adaptive feature recalibration of key regions and effectively suppress background noise. Additionally, a composite loss function consisting of mean squared error loss and perceptual loss is constructed to ensure pixel consistency while improving structural fidelity and visual naturalness. Comparative experiments were conducted based on the self-built CMUHL mine dataset, and the results show that the proposed algorithm achieves a PSNR of 30.544 dB and an SSIM of 0.914, outperforming advanced algorithms such as FLOL, AnlightenDiff, and Retinexformer. The model has only 4.37M parameters, effectively reducing computational complexity while maintaining enhancement performance. Real-scene verification results indicate that the algorithm achieves the best and second-best performance in the NIQE and BRISQUE no-reference metrics, with enhanced images featuring balanced brightness, clear details, and no obvious color bias, demonstrating good engineering application value. Ablation experiments further verify the synergistic benefits of RDB, CSA, and the composite loss function.