Title:Searching Efficient Model-guided Deep Network for Image Denoising

Abstract:Neural architecture search (NAS) has recently reshaped our understanding on various vision tasks. Similar to the success of NAS in high-level vision tasks, it is possible to find a memory and computationally efficient solution via NAS with highly competent denoising performance. However, the optimization gap between the super-network and the sub-architectures has remained an open issue in both low-level and high-level vision. In this paper, we present a novel approach to filling in this gap by connecting model-guided design with NAS (MoD-NAS) and demonstrate its application into image denoising. Specifically, we propose to construct a new search space under model-guided framework and develop more stable and efficient differential search strategies. MoD-NAS employs a highly reusable width search strategy and a densely connected search block to automatically select the operations of each layer as well as network width and depth via gradient descent. During the search process, the proposed MoG-NAS is capable of avoiding mode collapse due to the smoother search space designed under the model-guided framework. Experimental results on several popular datasets show that our MoD-NAS has achieved even better PSNR performance than current state-of-the-art methods with fewer parameters, lower number of flops, and less amount of testing time.