Unsupervised enhancement algorithm for real low-light mine images

LI Ran; WANG Haodian; MAO Zhihao

doi:10.11799/ce202412051

Abstract：In order to solve the problem of image degradation caused by insufficient light and dust interference in mines,and improve the visual perception of mine images and the accuracy of subsequent downstream tasks,an unsupervised enhancement method for low illuminance mine images is proposed.This method aims to overcome the limitations of traditional image processing techniques,which are limited by manually designed features and fixed processing flows,as well as the limitations of deep learning techniques,which are limited by the size and quality of data sets.A multi-scale structure-contrast enhancement network is designed.The network first decomposes the original input image into an image-Laplacian pyramid,and extracts rich multi-scale features through different scales of images.Next,the feature fusion module is used to implicitly embed high-frequency information into the enhancement process,thereby effectively enhancing the texture structure of the image while improving its contrast.In addition,this method uses unsupervised training methods,which do not rely on paired low-illuminance-normal-illuminance images,and only requires real images taken on-site in the mine for training,thus achieving improved generalization ability of the model.The experimental results show that the multi-scale structure-contrast enhancement network proposed in this paper achieves significant results in enhancing low-illuminance mine images.The image enhanced by this method has significantly improved in terms of contrast and texture structure,effectively improving the visual perception of the mine image and providing more accurate image information for subsequent downstream tasks.