Enhanced Multi-Attention Network for Single Image Super-resolution

Recent research on single image super-resolution(SISR) shows that deep convolutional neural networks(DCNNs) with attention mechanism present a better improvement. Each different attention mechanism has its distinct focus. Specifically, channel attention mechanism has the capacity to enhance the influence of critical channels by focusing on the expression of characteristics at different channel levels, and pixel attention mechanism has the ability to improve the quality of reconstructed images by paying attention to the expression of spatial pixel features. We believe that the combination of these two mechanisms is a way to further improve the quality of super-resolution image. In this paper, an enhanced multi-attention network(EMAN) is proposed, which contains advantages of two attention mechanisms. Besides, to improve the utilization of high-frequency information, a novel edge-based loss function is added to boost the learning of the edge region. Plenty of experiments show that the proposed multi-attention network achieves better accuracy and visual effect against single-attention methods.