Multi-scale joint network based on Retinex theory for low-light enhancement

Due to the limitations of devices, images taken in low-light environments are of low contrast and high noise without any manual intervention. Such images will affect the visual experience and hinder further visual processing tasks, such as target detection and target tracking. To alleviate this issue, we propose a multi-scale joint low-light enhancement network based on the Retinex theory. The network consists of a decomposition part and an enhancement part. As a joint network, the decomposition and enhancement parts are mutually constrained, and the parameters are updated at the same time so that the image processing results are more excellent in detail. Our algorithm avoids the separation and recombination of decomposition and enhancement. Therefore, less information is lost in the processing of low-light images, and the enhancement result of the proposed algorithm is very close to the ground truth. In addition, in the enhancement part, we adopt a multi-scale network to fully extract image features. The multi-scale network maintains a balance between the global and local luminance of the illumination image. Retinex theory can effectively solve the problem of noise amplification and color distortion. At the same time, we have added color loss to solve the problem of color distortion, so that the enhancement result is closer to the normal-light image in color. The enhancement results are intuitively excellent, and the peak signal-to-noise ratio and structural similarity index results also reflect the reliability of the algorithm.