Single image super-resolution via global aware external attention and multi-scale residual channel attention network

Recently, deep convolutional neural networks (CNNs) have shown significant advantages in improving the performance of single image super-resolution (SISR). To build an efficient network, multi-scale convolution is commonly incorporated into CNN-based SISR methods via scale features with different perceptive fields. However, the feature correlations of the same sample are not fully utilized by the existing multi-scale SISR approaches, impeding the further improvement of reconstruction performance. In addition, the correlations between different samples are still left unexplored. To address these problems, this paper proposes a deep-connected multi-scale residual attention network (DMRAN) by virtue of the feature correlations of the same sample and the correlations between different samples. Specifically, we propose a deep-connected multi-scale residual attention block (DMRAB) to take fully advantage of the multi-scale and hierarchical features, which can effectively learn the local interdependencies between channels by adjusting the channel features adaptively. Meanwhile, a global aware external attention (GAEA) is introduced to boost the performance of SISR by learning the correlations between all the samples. Furthermore, we develop a deep feature extraction structure (DFES), which seamlessly combines the stacked deep-connected multi-scale residual attention groups (DMRAG) with GAEA to learn deep feature representations incrementally. Extensive experimental results on the public benchmark datasets show the superiority of our DMRAN to the state-of-the-art SISR methods.