Research on Downhole Image Dehazing Algorithm Based on gUNet

Based on the complexity of the underground environment in mines, problems such as dust, fog, and dim lighting occur during coal mining operations, which affect the image quality captured by underground imaging equipment. Design an image dehazing algorithm suitable for underground mining to address the issues of edge and detail loss in dehazing networks. First, a Global Channel Spatial Attention (GCSA) module is designed based on gUNet, where spatial attention focuses on areas with higher fog density, and channel attention identifies color and texture feature channels that are sensitive to fog underground. Secondly, the introduction of the Multi-scale Convolutional Attention (MSCA) module improves the clarity of the underground passage profile and equipment edges. Then, the multi-scale dilated fusion attention (MDFA) module is designed to complete feature extraction of different sizes through convolutions with different dilation rates, enhancing the recognizability of distant targets and the restoration of near-sighted details under extreme lighting conditions. Finally, add the L1 loss function during training to optimize the network parameters. The results of the Haze4K dataset experiment show that the PSNR of the dehazed images based on this algorithm reaches 31.99 dB and the SSIM is 0.956. The PSNR obtained on the self-made underground dataset is 32.45 dB and the SSIM is 0.970.