Self-supervised multi-blind network for real image denoising via multivariate Gaussian-poisson noise

The noise in real images exhibits more complex distributions than the synthetic noise and distinguishes across different scenarios. Furthermore, the scarcity of "clean-to-noisy" paired image datasets makes the current models difficult to denoise successfully. To address these challenges, we propose MGP-MBF${\text{M}}^2$ANet, a self-supervised multi-blind feature multi-modulation attention network based on multivariate Gaussian-Poisson noise prior for real image denoising. Firstly, we propose a multivariate Gaussian-Poisson distribution to construct noisy images that contain more complex pixel spatial positions and intensity correlations, which expand the training domain and improve the model’s ability to generalize across diverse real noisy images. Building on this, we implement a random sampling mechanism based on four-neighborhood similarity to construct "noise-noise" training pairs, effectively exploiting the statistical properties of local structures in noisy images, without relying on any clean reference image. During the network design phase, a multi-blind feature multi-modulation attention module successfully enhances the representation of local features, which introduces multi-masked strategy to force network to learn more information to address the challenge of feature identity mapping. Experimental results demonstrate that the proposed method effectively suppresses noise and recovers high-frequency details within an unsupervised learning paradigm, achieving superior performance in both objective evaluation metrics and subjective visual quality across multiple real-world datasets.