A Retinex-Based Network for Low-Light Image Enhancement With Multi-Scale Denoising and Focal-Aware Reflections

Abstract

Low-light image enhancement addresses critical challenges in computer vision, including insufficient brightness, excessive noise, and loss of detail in low-light images, thus improving the quality and applicability of image data for various vision tasks. We propose an unsupervised Retinex-based network for low-light image enhancement, incorporating a multi-scale denoiser and focal-aware reflections. Our approach begins with a multi-scale denoising network that removes noise and redundant features while preserving both global and local image details. Subsequently, we employ an illumination separation network and a focal-aware reflection network to extract the illumination and reflection components, respectively. To enhance the accuracy of the reflection component and capture finer image details, we introduce a depthwise convolutional focal modulation block. This block improves the representative capacity of the reflection component feature map. Finally, we adjust the illumination component and synthesize it with the reflection component to generate the enhanced image. Extensive experiments conducted on 9 datasets and 13 methods, using metrics such as SSIM, PSNR, LPIPS, NIQE, and BRISQUE, demonstrate that the proposed method outperforms existing unsupervised approaches and shows competitive performance when compared to supervised methods.