FreqLight-Net: A Compact Frequency-Aware U-Net for Low-Light Image Enhancement

Abstract

Low-light image enhancement is a fundamental computer-vision problem that affects downstream tasks such as object detection, surveillance, autonomous driving, and night-time photography. Existing learning-based solutions either rely on heavy generative models with hundreds of millions of parameters or on purely spatial U-Net variants that ignore the strong frequency-domain priors associated with illumination and noise. The objective of this study is to design a compact yet competitive frequency-aware enhancement network. We evaluate the method on the LOL-v1 paired dataset (485 training, 15 testing image pairs) using standard PSNR, SSIM, and LPIPS metrics. The proposed FreqLight-Net is a five-million-parameter U-Net augmented with a Fast Fourier Transform (FFT) branch at the bottleneck, where spatial and spectral features are fused through a learnable convex combination before decoding. A composite loss combining L1 reconstruction, structural similarity, VGG-based perceptual, and frequency-spectrum losses guides the joint optimization of low-frequency illumination recovery and high-frequency detail preservation. Training is performed on an NVIDIA RTX-3080 GPU with the Adam optimizer, a cosine-annealed learning rate, and standard augmentation. Across ten independent runs, FreqLight-Net attains PSNR = 22.81 ± 0.64 dB, SSIM = 0.856 ± 0.004, and LPIPS = 0.298 ± 0.013 on LOL-v1, outperforming representative baselines including LIME, RetinexNet, Zero-DCE, MBLLEN, EnlightenGAN, and KinD while remaining substantially lighter than transformer-based competitors. This suggests that frequency-aware compact U-Nets offer a deployable path for real-time low-light enhancement on edge devices, and that combining spectral and spatial supervision is a generalizable principle for image restoration tasks.