Dynamic dual attention iterative network for image super-resolution

Recently, deep convolution neural networks (DCNNs) have obtained remarkable performance in exploring single image super-resolution (SISR). However, most of the existing CNN-based SISR methods only focus on increasing the width and depth of the network to improve SR performance, which makes them face a heavy computing burden. In this paper, we propose a lightweight dynamic dual attention iteration network (DDAIN) for SISR. Specifically, to better realize the attention of the channel and the convolution kernel, we design a dynamic convolution unit (DYCU) at the head of the network. It improves the SR performance by enhancing the complexity of the model without increasing the width and depth of the network. Compared with the traditional static convolution, it can extract more abundant high and low-frequency image features according to different input images. Moreover, to recover the high-frequency detail features of images with different resolutions as much as possible, we embed multiple dual residual attention (DRA) in the feature refinement unit (FRU). Finally, to alleviate the height discomfort caused by SR, we introduce iterative loss Liter to optimize the training process further. Extensive experimental results on benchmark show that the performance of the DDAIN in different degradation models exceeds some existing classical methods.