Single Image Super-Resolution using Adaptive Upsampling Convolutional Network
Peng Liu, Ying Hong, Yan Liu · 2019
These methods based on deep convolutional neural networks have achieved dramatic improvements in image super-resolution reconstruction. In this paper, we propose an adaptive upsampling convolutional network for single image super-resolution. The proposed network is mainly based on the adaptive upsampling convolutional unit (AUCU) which is composed of convolutional layers, parametric rectified linear units, sub-pixel convolution layer and the adaptive shortcuts. In the AUCU model, different weights will be assigned to different inputs of the adaptive shortcuts. And these weights are obtained adaptively from training. Owing to the special structure of AUCU, the proposed algorithm can recovery the fine texture details for a large upscaling factor. Besides the AUCU, bicubic interpolation algorithm is also used for the super-resolution restoration during the reconstruction process. In order to enhance the quality of the reconstructed image, a perceptual loss function is proposed for training the feed-forward networks. The proposed loss function consists of two parts of loss: feature loss and MSE loss. The experimental results on the pubic benchmark datasets demonstrate that the proposed algorithm outperforms many other state-of-the-art super-resolution methods.