Fast Single Image Super-Resolution Via Dilated Residual Networks

Abstract

Recently, deep convolutional neural networks (CNNs) have been attracting considerable attention in single image super-resolution. Some CNN-based methods, such as VDSR verified that residual learning can speed up the training and significantly improve the performance of accuracy. However, with very deep networks, convergence speed is still a critical issue in training due to the cost of requiring enormous parameters. In order to deal with this issue, we redesign the residual networks based on dilated networks. In this paper, we propose symmetrical dilated residual convolution networks (FDSR) to tackle image super-resolution problems. Our network is on the basis of the dilated convolutions supported exponential expansion of the receptive field without loss of resolution and coverage. This means that FDSR can speed up the training and improve the performance of accuracy without increasing the model's depth or complexity. Meanwhile, we attempt to combine the image pre-processing approach of VGG-net with mean squared error (MSE) to enhance the performance. The experimental results demonstrate that the training time-consuming proposed model achieves nearly a half with even superior restoration quality. Further, we present a novel network with less layers and parameters can achieve real-time performance on a generic CPU and still maintain superior performance. INDEX TERMS Image super-resolution, residual learning, dilated convolution.

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

“…They evaluated the superiority of HDC and achieved state-of-the-art result in semantic segmentation. Inspired by their work, Lu et al [43] further proved its effectiveness in SISR.…”

“…Compared with standard convolution, its biggest advantage is that it can expand the receptive fields by appropriately setting the rate of dilation without increasing more parameters. Recently, some algorithms [43], [44] have attempted to combine dilated convolution with super-resolution and have achieved noticeable performance. Whereas, the dilated convolution has its inherent problem that there exists ''gridding'' effect which causes losing a huge portion of contextual information [42].…”

“…The number of total layers is kept the same as in plain sub-network. Following the idea of [43], in order to increase the receptive field as much as possible and alleviate the ''gridding'' effect, the rate of dilation of each layer is set in a specific order (e.g., 1, 2, 3, 4, 3, 2, 1, respectively). The effective size of each filters is 3 × 3, which means no matter how much the dilated rate of a single kernel is, the pixels used for the computation is always 9.…”

“…In this part, several algorithms are used for quantitative comparison, including SRCNN [14], FSRCNN [16], VDSR [15], DRCN [20], CDCNN [44] and FDSR [43]. To make the results more convincing, we perform the evaluations on four widely used benchmark datasets with three sampling factors (×2, ×3 and ×4).…”

“…To make the results more convincing, we perform the evaluations on four widely used benchmark datasets with three sampling factors (×2, ×3 and ×4). Note that the results of CDCNN [44] and FDSR [43] are cited from their papers while the others are generated using the source codes publicly provided by the authors. Tab.…”