Transformer-Based Multi-scale Optimization Network for Low-Light Image Enhancement

Enhancing low-light images with high quality is a highly challenging task due to the features of low-light images such as brightness, color, and details in the information of different scales. Existing deep learning-based methods fail to fully utilize multi-scale features and fuse multi-scale features to comprehensively enhance the brightness, color and details of the images. To address these problems, a Transformer-based multi-scale optimization network for low-light image enhancement is proposed. Firstly, the Transformer-based multi-task enhancement module is designed. Through multi-task training, the Transformer-based enhancement module gains the ability to globally model brightness, color, and details. Therefore, it can initially cope with various degradation challenges commonly found in low-light images, such as insufficient brightness, color deviation, blurred details and severe noises. Then, the architecture combining global and local multi-scale features is designed to progressively optimize the features at different scales. Finally, a multi-scale feature fusion module and an adaptive enhancement module are proposed. They learn and fuse the information association among different scales, while adaptively enhancing images in various local multi-scale branches. Extensive experiments on six public datasets, including paired or unpaired images, show that the proposed method can effectively solve the problems of multiple degradation types, such as brightness, color, details and noise in low-light images. © 2023 Journal of Pattern Recognition and Artificial Intelligence. All rights reserved.