Content-aware Upsampling for Single Image Super-resolution

Super-resolution reconstruction has recently become a hotspot in digital image processing. It has been widely used in medical imaging, video surveillance security, high-definition television terminals, and other fields. In this paper, our group propose an image super-resolution method based on content-aware upsampling. In our neural network, the convolution kernel does not share parameters across the entire feature map, but the network generates specific convolution kernels at each pixel according to the content of the feature map. We use residual dense module as the backbone network. The following data shows that our method has more potential on training and testing datasets and the reconstructed image is more clearer than the sub-pixel convolution layer method.