RIE-CMU: reliable image enhancement method for low-light environments in underground coal mines

Abstract With the comprehensive development of intelligent coal mining, research into image enhancement techniques for underground coal mine (UCM) environments is gaining increasing popularity. This paper proposes a reliable image enhancement method, RIE-CMU, designed for low-light conditions in UCMs. It addresses issues such as poor dynamic range expansion, over-enhancement and colour distortion associated with the application of traditional Retinex-Net algorithms in UCMs, aiming to systematically resolve these problems. To address the challenge of feature extraction, we have designed a semantic embedding feature aggregation U-Net (SEFAU-Net) as the decomposition network. By incorporating techniques such as channel attention (CA) mechanisms , deformable convolutions, selective kernel feature fusion, and PixelShuffle pixel reorganisation, we have significantly enhanced feature extraction capabilities. Furthermore, we have designed a DenseUnet with convolutional architecture (DFC-Net) as the image fusion network. We fuse the results of contrast-limited adaptive histogram equalisation applied to low-light images with the reflection component. This optimises the image enhancement effect and reduces colour bias. By combining the improved decomposition network with the improved fusion network, we utilise the reflection and illumination components output by the decomposition network as inputs to the fusion network. Concurrently, we formulate a multi-task loss function with collaborative constraints to prevent error accumulation. This enhances the clarity and colour fidelity of the final image. The effectiveness of RIE-CMU has been validated through testing on the LOL dataset and our self-built UCM dataset, UCM. Experimental results demonstrate that, in terms of image enhancement quality, RIE-CMU achieves the highest peak signal-to-noise ratio (PSNR) and SSIM values compared to several benchmark methods, including KinD, Zero-DCE++, RRDNet, R2RNet, URetinex-Net++, and Retinex-Net. The PSNR and SSIM values of RIE-CMU are 22.43% and 70.81% higher, respectively, than those of the unmodified Retinex-Net. The learned perceptual image patch similarity and natural image quality evaluator values of RIE-CMU are reduced by 54.27% and 32.52%, respectively, compared to those of Retinex-Net. Additionally, ablation study and computational complexity analysis further validate the effectiveness of each module and the method’s adaptability to UCMs.