Multi-branch aware module with channel shuffle pixel-wise attention for lightweight image super-resolution

Deep convolutional neural networks (CNNs) have boosted the performance of image super-resolution (SR) in recent years. However, existing deep CNN-based SR approaches often have massive computation costs to obtain high performance, which are still difficult to be applied in real-world applications or resource-constrained devices. To address this problem, we propose a lightweight Multi-branch Aware Super-Resolution Network (MASRN) for cost-efficient image SR. Specifically, the multi-branch aware module (MAM) with a channel shuffle pixel-wise attention (CSPA) mechanism is proposed to constitute the basic build block of the network backbone. This lightweight module is also plug-and-play to extract hierarchical contextual features. Besides, we introduce a region-level (RL) feature fusion layer before the reconstruction stage of the network by utilizing the diversity of spatial positions to make feature representation non-locally enhanced. Extensive experiments on benchmark datasets demonstrate that the proposed MASRN achieves better accuracy and visual improvements against the state-of-the-art lightweight image SR methods in terms of qualitative and quantitative evaluation.