Adversarial noise generation modeling-based image denoising method for low-light CMOS sensors

Nowadays, image denoising technology for low-light CMOS sensors has been widely applied in many optical information processing-related applications, such as computational photography, security monitoring, and intelligent driving. Existing deep learning-based image denoising methods typically rely on large-scale, high-quality training datasets, with mainstream approaches using generated noisy low-light images to synthesize a large amount of paired training data. However, noise in low-light images exhibits complex statistical characteristics, especially signal-independent noise, making it difficult to achieve promising denoising results using traditional physical model-based noise generation methods. Besides, denoising networks trained on generated noisy images still have limitations in restoring image details and preserving texture information, which restrict their reconstruction accuracy and visual consistency in real low-light imaging environments. Therefore, in this paper, we propose an image denoising method for low-light CMOS sensors based on adversarial noise generation modeling. On the one hand, a frequency-domain-aware (FDA) adversarial noise generation model is constructed to learn and synthesize signal-independent noise with realistic statistical properties. In this model, an FDA discrimination mechanism is exploited to jointly constrain the generated noise from both spatial structure and frequency-domain distribution perspectives. On the other hand, a frequency-domain image denoising network, called FD-UNet, is specially designed. This network performs end-to-end modeling of the subband frequency-domain representation and introduces a channel attention mechanism in the frequency-domain enhancement branch to adaptively weight and adjust the feature channels. Besides, a residual reconstruction strategy is integrated into our FD-UNet in order to help preserve original image details while suppressing noise in an effective way. Finally, the noise generation and image denoising networks are jointly designed within a unified framework to construct a complete image denoising framework for low-light CMOS sensors. Experimental results on the public RLID and LOL datasets demonstrate that our proposed method achieves an average peak signal-to-noise ratio (PSNR) improvement of 0.34 d B and 1.36 d B , as well as an average structural similarity index (SSIM) improvement of 0.0004 and 0.0028 compared to the suboptimal method. Our code and dataset are available at https://github.com/MinjieWan/FD-UNet .