Lightweight image super-resolution reconstruction based on mixed attention and global inductive bias network

Convolutional neural network has significantly advanced the field of image super-resolution reconstruction in recent years. The insufficient ability to model global information of hierarchical features, incomplete attention to information, and excessive parameters make some based on convolutional methods face challenges. To address these issues, this paper designs the lightweight image super-resolution reconstruction based on Mixed Attention and Global Inductive bias Network(MAGIN), which directly captures the global information of phased features from the horizontal and vertical directions respectively to produce global inductive bias information. Furthermore, the global spatial attention block is proposed, which can realize the perception of global feature space without redundant operation. To realize more comprehensive attention for features, a mixed attention block arises, which enables the network to attend to features across different dimensions: channel, spatial, and pixel levels. Extensive experimental results demonstrate that the MAGIN delivers more satisfactory results than some other advanced lightweight reconstruction algorithms on several standard test datasets under the condition of considering the network’s performance, parameter and computational complexity.