Multi-attention augmented network for single image super-resolution

Highlights：

• We present a deep MAAN for high-quality image SR, which can jointly learn the optimal representation of multi-scale, multi-orientation and multi-level features. With the multi-stage strategy, our MAAN can obtain state-of-the-art results.

• We propose a gated U-net structure to generate the content-aware features, in which the spatial attention is developed to capture the long-range dependencies in different resolution feature maps. To further enhance feature discrimination, we design the combination of two pre-defined sparse kernels and one standard kernel to extract multi-orientation features, which are fused via channel attention mechanism.

• We propose a novel self-attention mechanism to help in recovering the realistic details. This attention is designed to refine the final feature maps according to the feature interaction in the neighbouring positions.