FRT-Net: A Lightweight Frequency-Spatial Framework for Low-Light Enhancement via Retinex Decomposition and FFT Filtering

Engineering notes

Experimental results on three mainstream benchmark datasets—LOLv1, LOLv2_real, and LOLv2_syn—demonstrate that FRT-Net ranks among the top two methods in terms of PSNR, SSIM, and LPIPS metrics, achieving an average PSNR of 22.98 dB, SSIM of 0.866, and LPIPS of only 0.075.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

In response to the challenges of low brightness, low contrast, and severe noise in low-light images, this paper proposes a lightweight, multi-scale, frequency- and spatial-domain collaborative low-light image enhancement network—FRT-Net. This method integrates classical Retinex theory with modern deep learning techniques. The Retinex decomposition module explicitly separates reflectance and illumination, providing the network with physical interpretability. A multi-scale feature extraction module is designed to capture global brightness trends and local texture details in parallel. Additionally, a frequency-domain FFT filtering branch is introduced to address spectral deficiencies and suppress noise. The CBAM attention mechanism is embedded to adaptively recalibrate channel and spatial weights, enhancing key feature representation. Finally, a comprehensive loss function is employed to collaboratively optimize brightness enhancement, detail recovery, color fidelity, and noise suppression. Experimental results on three mainstream benchmark datasets—LOLv1, LOLv2_real, and LOLv2_syn—demonstrate that FRT-Net ranks among the top two methods in terms of PSNR, SSIM, and LPIPS metrics, achieving an average PSNR of 22.98 dB, SSIM of 0.866, and LPIPS of only 0.075. The model contains only 0.55 million parameters(M) and requires 32.81 GFLOPs, meeting the real-time application demands of mobile devices. Ablation studies verify the effectiveness of each module. With its excellent performance and lightweight design, FRT-Net provides an efficient and robust visual perception foundation for practical applications such as nighttime autonomous driving and security surveillance.

Links and sources

Comments