Attention-guided multi-scale illumination estimation for low-light image enhancement

Abstract

Low-light image enhancement aims to solve the problems of low brightness and detail loss in images obtained under low illumination conditions. The design of decomposition and enhancement networks commonly used in this field mainly adopts simple convolutional neural network design, which lacks the guidance of relevant semantic information, resulting in certain noise and color distortion in enhanced images. To relieve the above problems, we designed a low-light image enhancement method based on attention mechanism and multi-scale illumination estimation. It can estimate illumination of low-light image at multiple global scales, and pay attention to the semantic information of spatial domain and channel domain through attention mechanism, so as to suppress the noise in the enhanced image. Then the detail reconstruction of the image after multi-scale illumination estimation fusion is carried out to reconstruct the detail and color of the image. Finally, we compared the current representative methods on the three evaluation criteria of PSNR, SSIM and LPIPS on commonly used public data sets. A large number of experimental results show that the proposed method has achieved better results in the subjective and objective comparison, and has excellent performance and potential.