基于光照引导的三阶段低光照图像增强

打开文本图片集

Three-stage low-light image enhancement based on illumination guidance

ZHANGYihang，ZHONGHan

（College of Information and Network Safety， People's Public Security University of China， Beijing100038，China） * Corresponding author， E -mail： zhonghan@ppsuc. edu. cn

Abstract： Addressing challenges such as severe noise，color deviation，and artifacts in low-light imaging， a novel three-stage low-light image enhancement algorithm based on illumination guidance （IG-TSNet） is proposed. This algorithm synergisticall integrates the Fourier domain's capability to capture global image information with the Transformer's strength in modeling long-range dependencies within the spatial domain，utilizing illumination guidance to ensure coherent enhancement. IG-TSNet comprises three sequential stages. In the pixel-wise enhancement stage，an adaptive parameter adjustment mechanism is introduced to improve the global representation of the image.During the Fourier reconstruction stage，illumination priors are employed to optimize both amplitude and phase spectra acrosstwo channels following Fouriertransformation，enabling comprehensive global image reconstruction. In the cross-atention fusion stage，a lightweight dual-path U-shaped network， incorporating a cross-attention fusion module，is designed to dynamically align Fourier-reconstructed features with illumination-guidance maps. The proposed IG-TSNet was rigorously evaluated on six benchmark datasets for low-light image enhancement，demonstrating superior performance.Qualitative results confirm that the method effectively enhances underexposed regions，suppresses noise without introducing artifacts or patchiness，and preserves color fidelity ro bustly.Quantitative assessments reveal that IG-TSNet achieves state-of-the-art results across nine evaluation metrics. On three paired datasets，PSNR values of 26.968 dB，27.880 dB，and 28.939 dB；SSIM values of O.867，0.882，and O.947;and LPIPS values of 0.099，0.141，and 0.047 were atined，respectively. On three unpaired datasets，BRISQUE scores of 25.67，20.51，and 18.80 and NIQE values of 3.79，4.O9，and 4.O2 were achieved，respectively. This study offers a viable frequency-spatial joint enhancement framework，advancing the field of low-light image enhancement.

Key words： image enhancement；low-light image;Fourier frequency domain；cross-attention；transformer

1引言

低光照条件是捕获图像数据时客观存在且较为常见的现实场景，在此情况下拍摄的图像不仅影响人眼观感，而且不利于其内部信息的准确获取和后续使用。（剩余22764字）