Retinex-SIE: self-supervised low-light image enhancement method based on Retinex and homomorphic filtering transformation

Abstract

Low-light images suffer from low visibility, much noise, uneven illumination distribution, etc. Many existing methods have problems such as over enhancement or insufficient detail enhancement when dealing with low-light images with uneven illumination distribution. To remedy the above shortcomings, we propose a Retinex-based self-supervised low-light image enhancement model (Retinex-SIE), which is mainly composed of three parts: Retinex-based self-supervised image decomposition network (Retinex-DNet), nonlinear conditional illumination enhancement function (NCIEF), and Image Reconstruction (IR). First, a uniform illumination image of the same scene with the low-light image is generated by homomorphic filtering transformation, and the low-light image and the uniform illumination image are input into Retinex-DNet for decomposition to obtain reflectivity, noise and illumination. Then, NCIEF is used to enhance the illumination after decomposition. Finally, the final enhanced image is obtained by multiplying the decomposed reflectance and the enhanced illumination. Experiments on severa challenging low-light image datasets show that Retinex-SIE proposed in this paper can better handle low-light images with uneven illumination distribution and avoid problems such as excessive enhancement or insufficient detail enhancement.