Attention-Plus-Plus Network for Lightweight Image Super-Resolution

Fig. 1. Parameters vs. PSNR vs. FLOPs on Manga109 dataset (×4). Fig. 1. Parameters vs. PSNR vs. FLOPs on Manga109 dataset (×4).Transformer-based image super-resolution (SR) algorithms have achieved remarkable progress due to their powerful long-range modeling capability. However, they still suffer from high computational complexity and insufficient multi-dimensional feature interaction, limiting the practical deployment in real-world scenarios. To address this issue, we propose a novel lightweight attention paradigm, dubbed Attention-Plus-Plus (A++) network. A++ first excavates self-attention information in channel dimension, and then exploits cross-attentions between those self-attention features and decomposed 2D spatial information along horizontal and vertical directions. Moreover, a multi-level feature extraction method is developed to deepen feature analyses in exploring local intrinsic characteristics. Furthermore, we design a multi-scale feed-forward block to achieve adaptive feature integration. This attention-cross-attention scheme strengthens multi-dimensional feature interactions while effectively reducing computational overhead. Experimental results demonstrate that the proposed A++ achieves superior performance-complexity tradeoffs compared to other state-of-the-art lightweight SR models, especially for the SR with large scales.