Single Image Super-Resolution with Hierarchical Receptive Field

As a pixel-level prediction task, it's crucial for single image super-resolution (SISR) to capture contextual information over the multi-scale regions in low-resolution (LR) space, which is used to predict the image details in high-resolution (HR) space. So researchers proposed multiple methods to enhance the size of receptive field and take the contextual information of images into account. But most of them tend to increase the depth of networks or the size of kernels simply, which is inefficient and consumes a large amount of computational resources and memory. In this paper, we combine dilated convolutions with standard convolutions and propose hierarchical receptive field network (HRFN) to enlarge receptive field without additional computational complexity. Specially, in each hierarchical receptive field block (HRFB), we apply standard convolutions with different kernel sizes and dilated convolutions with different dilation factors to adaptively obtain multi-scale features. Meanwhile, to ease the training process and make the model focus on the prediction of image details (high-frequency features), the residual learning is adopted locally and globally to explicitly learn and predict the difference between HR images and LR images. Finally, experimental results on five extensively used datasets show that our model outperforms those state-of-the-art methods for both quantitative and qualitative comparisons.