Multilevel Noise Contrastive Network for Few-Shot Image Denoising

In recent years, most denoising methods based on deep Convolutional Neural Networks heavily rely on massive noisy-clean image pairs. Collecting massive noisy-clean image pairs is expensive and not practical in real scenes. Currently, few-shot learning has been applied to many areas to cope with the absence data. The few-shot learning, however, in image denoising severely suffers from domain gap problems, including dataset domain gap and feature domain gap, especially for the real noisy images. Therefore, this paper proposes a Multi-level Noise Contrastive Network (MNC-Net) performing few-shot image denoising. MNC-Net consists of two training stages: i) using contrastive learning to self-supervise the training of Multi-level Noise Contrastive Learner (MNCL) on the pure synthetic noisy images with multiple Gaussian noise levels to ease the acute dataset domain gap, and ii) features generated by the MNCL on limited data are fused to the second stage and alleviate the feature domain gap using our proposed denoising network. Specifically, the MNCL consists of a Contrastive Feature Extractor (CFE) and a Contrastive Feature Projector (CFP). MNCL learns the rich and complex content-invariant degradations and general multiple-level noise representations. The denoising network in the second stage is composed of Guider Feature Encoder (GFE) and Adaptive Denoising Decoder (ADD). The GFE uses contrast features from CFE to guide the produced representations on the specific input noisy images. Then, such output features are fed into the ADD to adaptively denoise the noisy images on the corresponding noise distribution. To our best knowledge, this work is the first attempt to jointly use the few-shot learning and contrastive learning in the deep denoising field. Extensive experiments on CBSD68, Kodak24, Set12, SIDD, and DND show that our method achieves promising denoising performances in the absence of data.