UFR-GAN: A lightweight multi-degradation image restoration model

Abstract

Real-world pattern recognition systems often face challenges in simultaneously handling images degraded by multiple factors such as rain, haze, and low-light conditions. Existing single-degradation models fail to generalize across diverse scenarios, while specialized models are computationally expensive and inefficient. This paper introduces UFR-GAN, a lightweight multi-degradation restoration GAN-based framework. Our model integrates transformer-driven feature aggregation to enhance long-range dependencies while maintaining computational efficiency. Additionally, we employ frequency-domain contrastive learning to disentangle overlapping artifacts, thereby improving the restoration of fine details. This approach may enable unified restoration across diverse degradation types. Our UFR-GAN, which only had 45.4 M parameters, achieved state-of-the-art (SOTA) performance with 26.87 PSNR and 0.83 SSIM in various degradation scenarios. It also had significantly lower computational complexity than SOTA approaches. Crucially, we demonstrated its broader impact on downstream pattern recognition tasks: integrating UFR-GAN with YOLOv11 improved vehicle detection accuracy by 23% at 0.73 mAP50 while reducing the inference time to only $2.7ms$ under adverse weather conditions. These results indicated that our UFR-GAN could be used in many important areas of pattern recognition, including computer vision, remote sensing, traffic surveillance, and self-driving systems. Efficient and generalizable restoration techniques are essential in these domains.