Image Enhancement by Deep Learning Network Based on derived image and Retinex

Abstract

The Retinex model explains how the human visual system regulates the color and brightness of objects perceived by the human eyes. Based on this theory, an image enhancement algorithm based on image derived graph for weak illumination image by deep network is proposed in this paper. Firstly, in view of low contrast, low overall brightness and blurred details of dark areas in images captured under weak illumination, traditional image enhancement methods are used to process images to obtain derivative maps: adaptive contrast-limited histogram equalization, logarithmic histogram equalization, adaptive gamma correction with weighted distribution. Then, the image illumination component is obtained by deep decomposition network. The end-to-end mapping relationship between low illumination image and normal illumination image is trained, and the final output image is enhanced by deep enhancement network. The experimental results show that the proposed algorithm has better robustness as well as the output image has richer details, higher contrast, better visual effect and image quality compared with some state-of-the-arts.