Abstract: In order to solve the problems of low contrast and too dark brightness of images acquired in low-light scenes, an improved low-light image enhancement network based on Retinex is designed. Low-light images were decomposed into reflection maps and illumination maps in a data-driven way, noise suppression and detail restoration were performed on the reflection maps through UNet+KG-*4+, and adaptive brightness adjustment was achieved on the illumination maps according to the enhancement ratio set by the user. Validation on multiple datasets shows that our algorithm has certain advantages in vision, objective indicators, operation efficiency and can effectively enhance the low-light images under different light intensities. It has certain application value for the night image acquisition under the background of artificial intelligence.