An effective underground image enhancement approach based on improved KinD network

An effective underground image enhancement approach based on improved KinD network

연구 분야: Verification

학회: Signal, Image and Video Processing

Mine images often have problems such as unclear details and low miner-background contrast due to lighting issues, and traditional enhancement methods are difficult to effectively improve. To enhance the contrast and texture details of low-light mine images, this paper proposes an image enhancement method based on an improved KinD deep neural network. This method decomposes, adjusts, and reconstructs mine images by constructing an illumination decomposition network, an illumination adjustment network, and a reflection reconstruction network. The illumination decomposition network combines the multi-scale feature extraction ability of the Inception module with the residual connection mechanism of ResNet, enabling more accurate separation of illumination and reflection components from mine images, effectively avoiding complex lighting issues in mine images, and improving the accuracy and efficiency of image decomposition. The reflection reconstruction network introduces the MobileNetv3 network structure and replaces the SE attention mechanism with the CA attention mechanism, improving the network's feature learning ability and the restoration effect of reflection component texture details. Finally, the enhanced mine image is obtained by fusing the processed illumination and reflection components based on the Retinex theory. Experimental results show that the improved KinD method effectively improves the colorfulness, contrast, and clarity of the image, significantly improving the texture details of the image. This method has good effects in mine image enhancement.