Abstract: The complex environment of underground coal mining faces seriously affects the quality of monitoring images due to uneven suspended particles such as coal dust and water mist generated during operation. Common problems in image processing include incomplete dehazing, over-enhancement, and color distortion. The lack of paired hazy and clear images in underground environments limits the development of supervised dehazing models. To achieve effective dehazing processing and focus on the core issues of color fidelity and detail enhancement in post-dehazing images, this paper proposes a dual-branch fusion dehazing algorithm using unpaired data, consisting of four components: parameter estimation, color preservation branch, detail enhancement branch, and adaptive perceptual fusion. First, targeting the non-uniform distribution characteristics of hazy conditions in underground coal mines, a parameter estimation method based on hazy distribution features is employed to obtain initial transmission and atmospheric light values, which are substituted into the atmospheric scattering model to generate initial dehazed images. In the color preservation branch, a channel attention mechanism is embedded into a U-shaped network to optimize transmission rates. This processing improves dehazing effectiveness and avoids color distortion in non-dense hazy areas. The detail enhancement branch constructs a residual network architecture with dual attention mechanisms, achieving collaborative preservation of local details and global structural information while fusing spatial and channel information to enhance the expression capability of key features. Unpaired clear haze-free images are introduced for adversarial training to improve the detail representation of dehazed images. In the dual-branch fusion stage, an adaptive weighted fusion strategy is adopted that considers both color-sensitive regions and texture-complex regions to obtain final dehazing results. In the experiments, 876 images with 1920×1080 resolution were collected from coal mining faces (412 clear images and 464 hazy images), with 404 hazy images used for training and 60 for testing. Data augmentation was performed through cropping to 512×512 image patches and random horizontal and vertical flipping to ensure training sample diversity. The collected hazy images exhibit obvious non-uniform characteristics, including various complex distribution scenarios such as dense haze and thin haze, which are distinctly different from traditional uniform hazy images. The overall network significantly improves the model's generalization ability and robustness in non-uniform hazy environments while maintaining detail enhancement through the introduction of adversarial training strategies. Quantitative comparative experimental results show that the proposed method achieves an information entropy of 5.52, FADE reduced to 0.49, PIQE reduced to 7.39, and an average running time of 1.304s on the test set. The comprehensive performance is superior to various typical dehazing algorithms, effectively removing non-uniform haze, reducing color distortion, and improving image visualization effects. Ablation experimental analysis shows that: removing the color preservation branch leads to decreased image brightness, color distortion, and reduced visual quality; the absence of the detail enhancement branch causes texture blurring and local structural loss; removing the Squeeze-and-Excitation (SE) attention module reduces the discriminative power of key feature channels; eliminating the dual attention network results in severe haze residue and insufficient detail recovery. This indicates that the synergistic effect of each module plays a key role in improving the clarity, color consistency, and texture integrity of dehazed images. In conclusion, the dual-branch fusion dehazing network proposed in this paper can achieve efficient dehazing with color fidelity and detail enhancement in complex underground coal mine environments, balancing the removal of non-uniform haze with visual quality improvement, providing reliable support for the stable operation of on-site monitoring systems.