Deep Orthogonal Transform Feature for Image Denoising

Abstract

Recently, CNN-based image denoising has been investigated and shows better performance than conventional vision based techniques. However, there are still a couple of limits that are weak partly in restoring image details like textured regions or produce other artifacts. In this paper, we introduce noise-separable orthogonal transform features into a neural denoising framework. We specifically choose wavelet and PCA as an orthogonal transform, which achieved a good denoising performance conventionally. In addition to spatial image signals, the orthogonal transform features (OTFs) are fed into a denoising network. For the guide of the denoising process, we also concatenate OTFs from the image denoised by the existing method. This can play a role of prior for learning a denoising process. It has been confirmed that our proposed multi-input network can achieve better denoising performance than other single-input networks.