Clustered Adaptive Graph Priors Network for Few-shot Image Denoising

Recently, although current few-shot image denoising methods use limited samples to alleviate the problem of data scarcity, they still suffer from two main problems. Firstly, regular Convolutional Neural Networks (CNNs) cannot explicitly use the relationship between various parts of the image to construct global information in each layer, which hinders the image denoising performance promotion when using limited training samples. Secondly, due to limited training data, especially for noisy images with high diversity and complex noise, the retrieved features are insufficient to recover the denoised image. To address these issues, this work proposes a Clustered Adaptive Graph Priors Network (CAGP-Net) with few-shot learning for image denoising. Specifically, Graph Neural Networks (GNNs) are applied to capture local and global features. Then, a new method is proposed in CAGP-Net to transform image feature maps into graphs to generate local and global features and construct their interdependence. Additionally, to flexibly supplement the feature prior information required for image denoising, we propose a Clustering Adaptive Graph Prior (CAGP) mechanism to supplement the graph prior information learned from embedding features and provide adaptive graph prior information for restoring denoised images. Finally, to further improve the performance of few-shot image denoising methods, we propose a new two-stage few-shot learning strategy, which mainly aims to integrate the reconstruction feature maps from the first stage into each layer of CAGP-Net in the second stage to adaptively remove noise from noisy images. Extensive ablation and benchmark tests show that our CAGP-Net achieves state-of-the-art performance of the few-shot image denoising tasks.