Abstract: Abstact: The environment in underground coal mining faces complex challenges, where the operation generates a large amount of coal dust, water mist, and other unevenly distributed suspended particles, leading to significant degradation in the quality of monitoring images. Existing traditional algorithms suffer from poor dehazing effects, over-enhancement, and color distortion. Meanwhile, deep learning algorithms face the issue of lacking paired images of underground dust-mist and clear images. To address these problems, a dehazing algorithm based on a semi-supervised learning network is proposed. This semi-supervised learning network is composed of a generator and a discriminator: the generator adopts an encoder-decoder structure, where the encoder primarily uses a residual network as its main structure, incorporating a spatial attention mechanism in the residual blocks to better handle non-uniform dust and mist. The decoder consists of pixel shuffle layers and convolutional layers, progressively recovering higher-resolution feature maps. The discriminator outputs a probability map, representing the difference between the dehazed images generated by the generator and the real clear images. A contrastive learning branch is introduced to ensure that the dehazed images are closer to positive samples and farther from negative samples in the feature space, improving the model's generalization capability. Due to the lack of paired non-uniform dust-mist datasets in underground coal mining, a large number of images were collected from coal mine working faces. Based on the characteristics of dust and mist during operations, an atmospheric scattering model and Perlin noise were used to synthesize non-uniform dust-mist images on clear images. The synthetic data, along with the collected real data, were used to train the semi-supervised network, enhancing the model's adaptability and performance under non-uniform dust-mist conditions in underground coal mines. To validate the effectiveness of the proposed algorithm, four algorithms were selected for comparison. Experimental results show that the proposed algorithm effectively reduces the concentration of dust and mist in images, with minimal color distortion, thereby improving the visualization of the images.