HQPAFT: Enhancing Low-Light Images with High-Quality Priors and Advanced Feature Transformations Using Only Normal Light Images

Zeyu Li, Hanxiang Yang, Sheng Yang, Xiongxin Tang, Fanjiang Xu, Qiao Chen

Abstract: Improving low-light images is vital for fields like surveillance and medical imaging, where sharp and clear visuals are critical for accurate analysis. The primary challenge is the limited availability of high-quality data and the difficulty in enhancing image details without introducing errors. To address this challenge, we propose improving low-light enhancement through day-to-night domain adaptation. This approach aims to aid low-light scene enhancement by recovering well-lit scenes after degradation, eliminating the need for real low-light data. High-Quality Priors (HQPs) are introduced through discrete codebooks using a pre-trained VQGAN, Retinex-based reflectance decomposition, FFT-based feature transformation, and a bionic signal amplification mechanism. A comprehensive loss function ensures high-quality enhancement. Experimental results demonstrate that this method excels in noise reduction, detail preservation, and overall visual quality.