• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2018.06a
• /
• pp.63-66
• /
• 2018

Recently, a convolutional neural network (CNN) models at single image super-resolution have been very successful. Residual learning improves training stability and network performance in CNN. In this paper, we compare four convolutional neural network models for super-resolution (SR) to learn nonlinear mapping from low-resolution (LR) input image to high-resolution (HR) target image. Four models include general CNN model, global residual learning CNN model, local residual learning CNN model, and the CNN model with global and local residual learning. Experiment results show that the results are greatly affected by how skip connections are connected at the basic CNN network, and network trained with only global residual learning generates highest performance among four models at objective and subjective evaluations.

• International Journal of Computer Science & Network Security
• /
• v.21 no.11
• /
• pp.17-22
• /
• 2021

Recently, image super-resolution techniques used in convolutional neural networks (CNN) have led to remarkable performance in the research area of digital image processing applications and computer vision tasks. Convolutional layers stacked on top of each other can design a more complex network architecture, but they also use more memory in terms of the number of parameters and introduce the vanishing gradient problem during training. Furthermore, earlier approaches of single image super-resolution used interpolation technique as a pre-processing stage to upscale the low-resolution image into HR image. The design of these approaches is simple, but not effective and insert the newer unwanted pixels (noises) in the reconstructed HR image. In this paper, authors are propose a novel single image super-resolution architecture based on synchronized depthwise separable convolution with Dense Skip Connection Block (DSCB). In addition, unlike existing SR methods that only rely on single path, but our proposed method used the synchronizes path for generating the SISR image. Extensive quantitative and qualitative experiments show that our method (SDCN) achieves promising improvements than other state-of-the-art methods.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2023.05a
• /
• pp.121-121
• /
• 2023

다수의 서로 다른 해상도의 자료를 병합(Merge)하는 것은 강수 자료 사용에 중요한 절차 중 하나이다. 강수 자료는 다수의 소스(관측소, 레이더, 위성 등)에서 관측 자료를 제공한다. 연구자들은 각 원본 자료의 장점을 취하고 단점을 보완하기 위해 다중소스 기반의 재분석 강수 자료를 제작하여 사용하고 있다. 기존의 방법은 자료를 병합하기 위해 서로 다른 공간적 특성을 갖는 자료들을 공간적으로 동일한 위치로 보간(Interpolation) 하는 과정이 필요하다. 하지만 보간 절차는 원본자료에 인위적인 변형을 주기 때문에 많은 오차(Error)를 발생시키는 것으로 알려져 있다. 따라서 본 연구는 병합 과정에서 보간 절차를 제외하고 원본 해상도 자료를 그대로 입력하기 위해 머신 러닝 방법의 하나인 Super resolution convolutional neural network(SRCNN)에 기반한 병합 방법을 제안하고자 한다. 이 방법은 원본 자료의 영향을 모델이 직접 취사선택하여 최종 자료에 도달하기 때문에 병합 과정의 오류를 줄일 수 있을 것으로 기대된다.

• Journal of Information Processing Systems
• /
• v.18 no.1
• /
• pp.12-25
• /
• 2022

In recent years, deep convolutional neural networks have made significant progress in the research of single image super-resolution. However, it is difficult to be applied in practical computing terminals or embedded devices due to a large number of parameters and computational effort. To balance these problems, we propose CSRNet, a lightweight neural network based on channel split residual learning structure, to reconstruct highresolution images from low-resolution images. Lightweight refers to designing a neural network with fewer parameters and a simplified structure for lower memory consumption and faster inference speed. At the same time, it is ensured that the performance of recovering high-resolution images is not degraded. In CSRNet, we reduce the parameters and computation by channel split residual learning. Simultaneously, we propose a double-upsampling network structure to improve the performance of the lightweight super-resolution network and make it easy to train. Finally, we propose a new evaluation metric for the lightweight approaches named 100_FPS. Experiments show that our proposed CSRNet not only speeds up the inference of the neural network and reduces memory consumption, but also performs well on single image super-resolution.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.06a
• /
• pp.98-101
• /
• 2019

At present, deep convolutional neural networks have made a very important contribution in single-image super-resolution. Through the learning of the neural networks, the features of input images are transformed and combined to establish a nonlinear mapping of low-resolution images to high-resolution images. Some previous methods are difficult to train and take up a lot of memory. In this paper, we proposed a simple and compact deep recursive residual network learning the features for single image super resolution. Global residual learning and local residual learning are used to reduce the problems of training deep neural networks. And the recursive structure controls the number of parameters to save memory. Experimental results show that the proposed method improved image qualities that occur in previous methods.

• Journal of Korea Multimedia Society
• /
• v.20 no.11
• /
• pp.1750-1758
• /
• 2017

In super-resolution, various methods with Convolutional Neural Network(CNN) have recently been proposed. CNN based methods provide much higher image quality than conventional methods. Especially, VDSR outperforms other CNN based methods in terms of image quality. However, it requires a high computational complexity which prevents real-time processing. In this paper, the method to apply a deconvolution layer to VDSR is proposed to reduce computational complexity. Compared to original VDSR, the proposed method achieves the 4.46 times speed-up and its degradation in image quality is less than -0.1 dB which is negligible.

• Journal of the Society of Naval Architects of Korea
• /
• v.59 no.2
• /
• pp.72-79
• /
• 2022

Deep-learning-based Super-Resolution (SR) methods were evaluated to reconstruct pressure fields with a high resolution from low-resolution images taken from a coarse grid simulation. In addition to a canonical SRCNN(super-resolution convolutional neural network) model, two modified models from SRCNN, adding an activation function (ReLU or Sigmoid function) to the output layer, were considered in the present study. High resolution images obtained by three models were more vivid and reliable qualitatively, compared with a conventional super-resolution method of bicubic interpolation. A quantitative comparison of statistical similarity showed that SRCNN model with Sigmoid function achieved best performance with less dependency on original resolution of input images.

• Journal of Multimedia Information System
• /
• v.8 no.2
• /
• pp.101-110
• /
• 2021

In computer vision, single-image super resolution has been an area of research for a significant period. Traditional techniques involve interpolation-based methods such as Nearest-neighbor, Bilinear, and Bicubic for image restoration. Although implementations of convolutional neural networks have provided outstanding results in recent years, efficiency and single model multi-scalability have been its challenges. Furthermore, previous works haven't placed enough emphasis on real-number scalability. Interpolation-based techniques, however, have no limit in terms of scalability as they are able to upscale images to any desired size. In this paper, we propose a convolutional neural network possessing the advantages of the interpolation-based techniques, which is also efficient, deeming it suitable in practical implementations. It consists of convolutional layers applied on the low-resolution space, post-up-sampling along the end hidden layers, and additional layers on high-resolution space. Up-sampling is applied on a multiple channeled feature map via bicubic interpolation using a single model. Experiments on architectural structure, layer reduction, and real-number scale training are executed with results proving efficient amongst multi-scale learning (including scale multi-path-learning) based models.

• Journal of Internet Computing and Services
• /
• v.19 no.4
• /
• pp.45-52
• /
• 2018

A single-image super-resolution is a process of restoring a high-resolution image from a low-resolution image. Recently, the super-resolution using the deep neural network has shown good results. In this paper, we propose a neural network structure that improves speed and performance over conventional neural network based super-resolution methods. To do this, we analyze the conventional neural network based super-resolution methods and propose solutions. The proposed method reduce the 5 stages of the conventional method to 3 stages. Then we have studied the optimal width and depth by experimenting on the width and depth of the network. Experimental results have shown that the proposed method improves the disadvantages of the conventional methods. The proposed neural network structure showed superior performance and speed than the conventional method.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.12
• /
• pp.1604-1611
• /
• 2020

Light field image captured by a microlens array-based camera has many limitations in practical use due to its low spatial resolution and angular resolution. High spatial resolution images can be easily acquired with a single image super-resolution technique that has been studied a lot recently. But there is a problem in that high angular resolution images are distorted in the process of using disparity information inherent among images, and thus it is difficult to obtain a high-quality angular resolution image. In this paper, we propose light field angular super-resolution that extracts an initial feature map using an dilated convolutional neural network in order to effectively extract the view difference information inherent among images and generates target image using a residual neural network. The proposed network showed superior performance in PSNR and subjective image quality compared to existing angular super-resolution networks.