Improving Super Resolution Methods Via Incremental Residual Learning

Abstract

Recently, Convolutional Neural Networks (CNNs) have shown promising performance in super-resolution (SR). However, these methods operate primarily on Low Resolution (LR) inputs for memory efficiency but this limits, as we demonstrate, their ability to (i) model high frequency information; and (ii) smoothly translate from LR to High Resolution (HR) space. To this end, we propose a novel Incremental Residual Learning (IRL) framework to address these mentioned issues. In IRL, first we select a typical SR pre-trained network as a master branch. Next we sequentially train and add residual branches to the main branch, where each residual branch is learned to model accumulated residuals of all previous branches. We plug state of the art methods in IRL framework and demonstrate consistent performance improvement on public benchmark datasets to set a new state of the art for SR at only ≈ 20% increase in training time. 1

Single-image super-resolution reconstruction using dark channel regularization network

“…In order to generate more photo-realistic results, Ledig et al further developed a GAN-based network, the SRGAN, which is a combination of SRResNet, the VGG-based perceptual loss, and a discriminator. Based on the success of SRResNet and SRGAN, many variants have been developed so far, including EDSR [20], RCAN [21], ASRResNet [8], IRL [22], and WDSRGAN [23]. EDSR eliminated the batch normalization layers and expanded the model size of SRResNet.…”

“…Among the NN-based methods, references [8,[14][15][16][17][18][20][21][22] try to solve Eq. (1) and no image prior is utilized.…”