PRNet: Low-Light Image Enhancement Based on Fourier Transform

Abstract

Low-light image enhancement (LLIE) techniques constitute a significant approach for enhancing image brightness effectively while preserving image details. In this article, PRNet is proposed, which is a novel lightweight LLIE network that leverages the Fourier transform, performing LLIE in two stages. In the first stage, a pixel enhancement network (PENet) enhances the brightness of the low-light image (LLI) through a dense skip-connection structure. This structure incorporates a custom-designed Fourier-based brightness enhancement block (FBEB). In the second stage, a refinement and restoration network (RRNet) processes the output from the first stage, further restoring image details. Detailed refinement is achieved using a dual-branch UNet structure, incorporating a bidirectional frequency-domain cross-attention solver (BFDCS) to optimize image quality. To thoroughly assess the performance of the proposed PRNet, nine well-established benchmark datasets were employed for detailed quantitative and qualitative evaluations. The experimental results show that PRNet achieves high-quality image enhancement at significantly reduced computational complexity.