The Modified Unsupervised Low-Light Image Enhancement Approach Based on the Retinex Theory

Low-light image enhancement technology is mainly used to process the images which are taken under the conditions of short exposure time of the device or dark ambient light. Aiming at the effective enhancement of the brightness of images and the preservation of the details as well, a light enhancement algorithm based on Retinex is proposed. First of all, a decomposition network is built based on Convolutional Neural Networks (CNN). The input is the low-illumination image with its maximum channel. Attention modules in space and channel dimensions are used to improve the capability of feature extraction. Then, a block adaptive illumination enhancement function is proposed to enhance the light of the divided image blocks, which can significantly improve the light of extremely low-light images. The loss function consists of three parts: reconstruction loss, reflection loss and illumination loss. The proposed loss function can effectively retain image details. The experiment demonstrates that the algorithm has an ideal result in low illumination image enhancement. The details of the enhanced image are clear, the noise suppression effect is better and the color distortion is smaller. Our approach is better in terms of subjective effects and objective evaluation indicators.