Efficient Low Light Video Enhancement Based on Improved Retinex Algorithms

Abstract

Videos shot in low-light environments suffer from low contrast and high noise. In this paper, an improved zero-reference low-light enhancement technique for videos based on the Retinex model is presented. The proposed method improves the existing Retinex approaches in several aspects. First, the image features extracted by the VGG network are employed as a part of the input to the generator of the Retinex parameters for increasing temporal stability. Second, a deformable convolution kernel is used to enhance the spatial correlation. Third, the optical flow between frames is approximated as a combination of affine linear transformations for reducing complexity. Compared with the state-of-the-art low-light enhancement algorithms, the proposed method achieves more favorable and stable image qualities in PSNR and SSIM with short processing time.