Low-light Image Enhancement Network Based on Multichannel Retinex Model

Low-light image enhancement has been one of the hottest research fields of computer vision in recent years. It has many applications in object detection, autonomous driving, and night monitoring. The pixel value distribution of images obtained from the same scene is analyzed under different exposures. It finds differences in the growth ratio of its low-light and normal-illumination images in RGB three channels. Based on this observation, a low-light image enhancement network is proposed on the basis of multi-channel Retinex model. In order to solve the problem of color deviation after low-light enhancement, a multi-channel enhancement strategy is adopted in the light enhancement module, and a targeted color loss function is designed, which improves the quality of generated pictures through the antagonistic loss function. Experimental results show that the peak signal-to-noise ratio between the enhanced image and the reference image is improved by 20% by the proposed method in comparison with the existing advanced algorithms through experiments on two public datasets, and structural similarity is improved by 7.2%. The noise of image is eliminated, and it is closer to the reference image in terms of numerical indicators and visual effects.