Differentiable Neural Architecture Search Based on Efficient Architecture for Lightweight Image Super-Resolution

With the advancement of deep neural networks, image Super-Resolution (SR) has witnessed remarkable improvements in performance. However, the increasing number of parameters and computational complexity has posed challenges for the practical deployment of SR models. To address these challenges, we propose a novel approach called Differentiable Neural Architecture Search (NAS) based on Efficient Architecture for lightweight image Super-Resolution, referred to as DNAS-EASR. In DNAS-EASR, we employ the information distillation mechanism (IDM) at the cell-level space to search for key operations. Additionally, we search for attention modules at the cell-level space to determine the most suitable attention module for our architecture. Furthermore, we adopt a hierarchical architecture as our backbone network to enable multi-scale information processing and fusion. Extensive experiments conducted on benchmark datasets demonstrate that DNAS-EASR is lightweight, efficient and capable of achieving comparable performance to other lightweight methods.