A Self-regularized Low-light Image Enhancement Based on the Brightness Information

Abstract

Low-light images have low brightness, low contrast, narrow gray range, color distortion and other problems, resulting in poor visual perception of the image, but also lead to subsequent image recognition, classification and other tasks of the accuracy of the reduction. However, existing low-light image enhancement algorithm has the difficulty of obtaining the training pairing data, and the color recovery of the enhanced image has some problems. In view of the above problems, we propose a self-regularized lowlight image enhancement algorithm based on the brightness information of HSV color space. First it converts low-light images to HSV space, preserves color features (hue, saturation) in lowlight images. Second, builds a brightness enhancement network on the value channel to enhance the brightness characteristics of lowlight images. The brightness enhancement network does not need paired low-light and normal lighting images, but uses non-reference loss function to regularize the network, effectively avoids the problem of over-fitting the label data with the model. Only the single channel of low-light images is enhanced, the computational complexity of the network is low and the computing speed is fast. Experiments show that our method can effectively enhance the low-light image, and can well retain the original color characteristics of the image, reduce the color deviation of the enhanced image, in the objective index and the sense of the supervisor is better than the current multiple method.