Abstract:
The existing methods for defogging coal mine images fail to perform lighting correction while extracting deep level feature information, resulting in the loss of detail information or image darkening in the processed images. A defogging algorithm for coal mine underground images based on dark channel guided filtering and lighting correction is proposed. Firstly, the original underground images are subjected to bilateral filtering, lighting estimation, and dark primary color processing using the image differentiation module (IDM) to obtain lighting maps, dark primary color maps, and lighting reflection maps. Secondly, the method preprocesses the dark primary color map and uses it as a weight guidance parameter to guide the filtering of the lighting reflection map, in order to restore the image's detailed feature information. Finally, the lighting map is used as a weight parameter to perform lighting correction and feature extraction on the image. The color distortion problem is solved through multiple lighting corrections, while increasing the network depth to remove degradation in dark areas, achieving reconstruction of image details and obtaining clear images. The subjective evaluation results indicate that the coal mine underground image defogging algorithm based on dark channel guided filtering and lighting correction retains more structural textures and background details while removing fog. It makes the entire image closer to the corresponding clear image. The objective evaluation results show that compared with the suboptimal algorithm PMS-Net, the information entropy on the training and testing sets is increased by 0.32 and 0.11, the standard deviation is increased by 3.58 and 1.89, and the average gradient is increased by 0.008 and 0.004, respectively. This indicates that the proposed algorithm can effectively reduce the fog in coal mine underground images. The results of ablation experiments show that the proposed algorithm has higher information entropy, standard deviation, and average gradient on the test dataset than other network models. It indicates that the defogging effect is the best and it can effectively preserve image details and edge information.