Single image super-resolution based on multi-scale dense attention network

Deep convolutional neural networks (CNNs) have enabled significant progress in the single image super-resolution (SISR) field in recent years. However, most super-resolution methods based on deep neural networks obtain spatial information using a single-size convolution kernel, which leads to insufficient local feature extraction. In addition, since convolution serves for local operations, it fails to capture the image's inherent attributes. This paper tackles these problems by proposing a novel multi-scale dense attention network (MSDAN) based on the attention mechanism and multi-scale feature extraction. The network employs multi-scale dense blocks (MSDB) that utilize convolution kernels of different sizes to extract feature information of distinct scales. Additionally, the attention is used to model global features effectively and strengthen the network's expressiveness. Hierarchical feature fusion is followed by multi-level feature extraction, which fuses each unit's feature map output. Finally, the image is reconstructed using up-sampling. The experimental results show that the proposed network outperforms other several state-of-the-art super-resolution reconstruction algorithms regarding objective evaluation indicators and visual effects on datasets Set5, Set14, BSD100 and Urban100.