Underground low-light self-supervised image enhancement method based on structure and texture perception

PAN Shan; YU Ting; CHEN Wei; TIAN Zijian; YUE Zhongwen

doi:10.13225/j.cnki.jccs.2024.0234

Abstract：Due to the complex spatial environment and uneven artificial lighting underground,images captured by under-ground visual equipment often suffer from insufficient overall or partial lighting and poor visibility of image content.Ex-isting image enhancement methods for low-light underground images often result in poor contrast and issues with overex-posure and underexposure in certain areas.Within this article,we propose a self-supervised image enhancement method for low-light underground conditions based on structural and texture perception,aiming to alleviate the dependence on paired low-light/normal-light images during training.Firstly,to generate smoothly segmented illumination maps,we design a self-supervised structural and texture-aware illumination estimation network,which preserves scene edge struc-tures and smooths texture details based on self-supervised training losses.To further exploit local texture features and global structural features in low-light images to improve the performance of the illumination estimation network,we intro-duce a local-global perception module into the illumination estimation network.This module leverages the ability of con-volutional operations with small receptive fields to capture local features and the self-attention mechanism of visual trans-formers to facilitate global information interaction,thus extracting local and global features from low-light images.Secondly,to facilitate self-supervised learning,we adopt a structure-aware smoothness loss considering the segmented smoothness property of illumination maps.Finally,to refine the illumination maps generated by the illumination estima-tion network for reasonable brightness and contrast,we introduce a pseudo-label image generator to synthesize pseudo-la-bel images with good contrast and brightness.By constraining the consistency between brightness and contrast of the en-hanced images and pseudo-label images through reconstruction loss,we indirectly constrain the illumination maps.Experi-mental results on multiple public benchmark datasets and a dataset of low-light images in real underground scenes demon-strate the effectiveness of our method in enhancing low-light images,as well as its good generalization performance when faced with low-light images in underground scenarios.