Lightweight image super-resolution network fusing multi-attention mechanism and blueprint separable convolution

Abstract

Aiming at the problems of large parameters and high computational complexity in deep learning-based image super-resolution networks, this paper proposes a lightweight super-resolution network that fuses multi-attention mechanism and Blueprint Separable Convolution (BSConv). BSConv is introduced to improve performance while reducing parameters. A Window-Shifted Spatial Self-Attention Module (WSSM) and a Blueprint Separable Channel-self-attention Module (BSCM) are designed to extract local features. A Coordinate Attention Module (CAM) is combined to capture global information. Extensive experiments on Set5, Set14, B100, and Urban100 datasets for ×2/×3/×4 reconstruction show that compared with the lightweight NGswin model, the proposed method improves PSNR by 0.09 dB on Set5 for ×2 reconstruction and reduces parameters by 453K, which can recover image textures and details more efficiently.