Abstract: Coal mining operations in underground coal mine working faces generate non-uniform suspended particles such as coal dust and water mist, causing video images to suffer from degradation phenomena including blurring and detail loss. Existing underground image enhancement algorithms demonstrate limited effectiveness, suffering from image distortion and poor generalization capabilities. To address these issues, this paper proposes a semi-supervised dual-branch network dehazing algorithm. The proposed algorithm employs an end-to-end deep learning network to map degraded images to clear images, trained via a semi-supervised learning paradigm. The network architecture comprises a dehazing branch and a detail restoration branch: the dehazing branch follows an encoder-decoder structure, where the encoder utilizes central difference convolution as the feature extractor to enhance the learning and representation of high-frequency information, and the decoder incorporates a denoising module to strengthen the dehazing capability; the detail restoration branch is constructed with DCResBlock modules, leveraging dilated convolutions to enlarge the receptive field and introducing minimal parameters to address detail loss during the dehazing process. Due to the difficulty in collecting paired datasets from coal mining working faces, this study improves upon existing data synthesis methods based on the characteristics of coal mine environments, synthesizing clear images into non-uniform dust and haze images that serve as training samples alongside real dust and haze images for semi-supervised network training. To mitigate the issue of poor algorithm generalization, contrastive learning loss is introduced as an unsupervised loss to guide model optimization and enhance dehazing performance on real dust and haze images. To validate the effectiveness of the proposed algorithm, five dehazing algorithms—atmospheric scattering model-based method, YOLY, AECR-Net, CasDyF-Net, and SSID—are selected for comparative experiments on a coal mine working face dust and haze image dataset. Experimental results demonstrate that the proposed algorithm can more effectively reduce dust and haze density, restore detail information, improve the visual quality and image quality of underground coal mining working face scenes, and enhance its practicality for engineering applications.