Computationally Efficient Neural Architecture Search for Image Denoising

In image denoising, deep neural networks require suitable architectural selection and noise-representative datasets for learning, often leading to significant computational demand. In this paper, this challenge is addressed in two ways. First, deep reinforcement learning is leveraged with weight-sharing for optimal architecture search, creating a shared model that enables efficient weight reuse across architectural configurations. This weight-sharing mechanism significantly reduces memory requirements. Second, by utilizing a representative subset of a dataset and examining empirical distributions of SSIM and PSNR values, it is demonstrated that comparable performance is achievable while reducing computational costs, provided that the subset is sufficiently sampled as measured by the Jensen-Shannon Divergence. This sufficiently sampled subset thus captures the characteristics of the full dataset. This combined approach of intelligent sampling and weight-shared reinforcement learning yields results equivalent to full-dataset training but with significantly reduced training time, demonstrating that efficient neural architecture search coupled with smart sampling can maintain high-quality denoising performance while substantially decreasing computational demand.