Single Image Super-Resolution Using Dual-Branch Convolutional Neural Network

Abstract

Recent advances in convolutional neural networks have demonstrated impressive reconstruction for single image super-resolution (SR). However, most of the existing methods are to achieve better performance by directly stacking more convolution layers in a chain way. As the depth increases, it not only lacks time efficiency but also consumes more computer memory. In this paper, we propose a lowcomplexity dual-branch network, termed DBCN, to deal with the SR problem. The feature extraction part of the proposed DBCN mainly consists of two parallel sub-networks, one extracts fine local information via standard convolution layers, and the other focuses on more contextual information in larger regions by using dilated convolution layers. In addition, deconvolution layers are integrated into the network after feature extraction stage to accelerate the upscaling process. Specifically, a novel skip connection is proposed to learn residual mapping and ease the training. Extensive evaluations on benchmark datasets show the effectiveness of our algorithm, achieving comparable results in terms of accuracy and visual quality compared with several state-of-the-art methods.

Deep Learning Algorithms for Single Image Super-Resolution: A Systematic Review

“…Environment, flora, fauna, handmade objects, people, and scenery are the key items found in the dataset [36]. Enhanced deep residual network (EDSR) [37], multi-connected convolutional network for super-resolution (MCSR) [38], cascading residual network (CRN) [39], enhanced residual network (ERN) [39], residual dense network (RDN) [40], Dilated-RDN [35], dense space attention network (DSAN) [41] and dual-branch convolutional neural network (DBCN) [42] were the algorithms that used the DIV2K dataset as their training dataset.…”

“…This has increased the running time and memory complexity of the model. Therefore, a dual branch-based image super-resolution algorithm was proposed by Gao et al [42], named DBCN. Figure 19 showed the network design of a DBCN.…”

“…From these aspects mentioned, DBCN improved at a rate of 0.68% compared to DRCN, whereas 0.74% improvement was made compared to VDSR. lution algorithm was proposed by Gao et al [42], named DBCN. Figure 19 showed the network design of a DBCN.…”

Multi-Perspective Discriminators-Based Generative Adversarial Network for Image Super Resolution

“…They show less blurred results than MSE-based deep learning SR, but they also have side effects such as the occurrence of artifacts and lower PSNRs. Although novel network structures for deep learning SR have been proposed [42]- [53], they all still have certain limitations.…”

“…As interest in deep learning has recently increased, a few deep learning-based SR methods have been proposed [9]- [16], [42]- [53]. Deep networks allow us to easily…”