From Local to Global: Efficient Dual Attention Mechanism for Single Image Super-Resolution

Convolutional neural networks (CNNs) have become a powerful approach for single image super-resolution (SISR). Recently, attention mechanisms are incorporated to enhance the network performance further. However, most methods use them locally to gather and model information at a single layer, which is not sufficient to capture the hierarchical relationship among various channels and restore high-frequency features. Here, we propose an efficient dual attention mechanism, with a global cross-layer attention (GCA) mechanism to emphasize high-frequency information learning by modeling cross-layer feature dependencies, and a local enhanced attention (LEA) mechanism, complementing GCA by offering attention-aware features for accurate feature fusion and also facilitating structure preservation. Experiments demonstrate that our method adapts well to multiple image degradation models and performs favorably against state-of-the-art methods.