This repository is an official implementation of the ECCVW 2020 paper "Multi-Attention Based Ultra Lightweight Image Super-Resolution".
Abstract: Lightweight image super-resolution (SR) networks have the utmost significance for real-world applications. There are several deep learning based SR methods with remarkable performance, but their memory and computational cost are hindrances in practical usage. To tackle this problem, we propose a Multi-Attentive Feature Fusion Super-Resolution Network (MAFFSRN). MAFFSRN consists of proposed feature fusion groups (FFGs) that serve as a feature extraction block. Each FFG contains a stack of proposed multi-attention blocks (MAB) that are combined in a novel feature fusion structure. Further, the MAB with a cost-efficient attention mechanism (CEA) helps us to refine and extract the features using multiple attention mechanisms. The comprehensive experiments show the superiority of our model over the existing state-of-the-art. We participated in AIM 2020 efficient SR challenge with our MAFFSRN model and won 1st, 3rd, and 4th places in memory usage, floating-point operations (FLOPs) and number of parameters, respectively.