Feature enhancement network architecture for low-light image

Abstract Limitations such as shooting equipment and environmental conditions can significantly impact the quality of captured images. Image enhancement, a crucial task in computer vision, aims to improve the overall or local characteristics of an image, particularly its brightness and contrast, to enhance the visual effect. The Retinex model has proven effective for low-light image enhancement. In this work, we propose a novel feature enhancement network architecture that combines the Retinex model with deep learning. The network comprises three main components: a decomposition network for image decomposition, an enhancement network for the luminance component obtained from decomposition, and a denoising network for the reflection component. The network’s learning process relies primarily on key constraints, including the consistency of the reflection component and the smoothness of the luminance component post-image decomposition. We conducted experiments on synthetic datasets involving real environments and processed images, and the results demonstrate that our method strikes a good balance between reducing parameters and maintaining high-quality image enhancement compared to other algorithms.