A new image decomposition model based on DIP and RED

Image decomposition is an important problem in image processing, which aims at decomposing a given image into structural components and oscillatory components. Although existing learning-based methods can get better structural features, they require a large number of samples for training, which are not present in cartoon-texture decomposition problems. Deep image prior (DIP) is a typical unsupervised deep learning method, which avoids collecting a large number of training samples. In this paper, we aim to boost DIP by adding an explicit prior of Regularization by denoising (RED), which can enrich the overall regularization effect in order to lead to better decomposition results. Our work shows how DIP and RED can be merged into a image decomposition model, and how the new model is solved efficiently. Finally, the experimental results are reported to show the visual qualities compared with some state-of-the-art methods.