Deep unfolding multi-scale regularizer network for image denoising

Abstract Existing deep unfolding methods unroll an optimization algorithm with a fixed number of steps, and utilize convolutional neural networks (CNNs) to learn data-driven priors. However, their performance is limited for two main reasons. Firstly, priors learned in feature space need be converted the image at each iteration step, which limits depth CNNs prevents from exploiting contextual information. Secondly, existing only single full-resolution scale, so ignore benefits multi-scale context dealing high level noise. To address these issues, we explicitly consider denoising process propose regularizer network (DUMRN) denoising. The core DUMRN feature-based module (FDM) that directly removes noise space. In FDM, construct block prior information multi-resolution features. We build by stacking sequence FDMs train it end-to-end manner. Experimental results on synthetic real-world benchmarks demonstrate performs favorably compared state-of-the-art methods.