• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.06a
• /
• pp.221-223
• /
• 2022

최근 CNN(Convolutional Neural Network)은 초해상화(super-resolution)를 포함한 다양한 컴퓨터 비전 분야에서 우수한 성능을 보이며 널리 사용되고 있다. 그러나 CNN은 계산 집약적이고 많은 메모리가 요구되어 한정적인 하드웨어 자원인 모바일이나 IoT(Internet of Things) 기기에 적용하기 어렵다는 문제가 있다. 이런 한계를 해결하기 위해, 기 학습된 깊은 CNN 모델의 성능을 최대한 유지하며 네트워크의 깊이나 크기를 줄이는 경량화 연구가 활발히 진행되고 있다. 본 논문은 네트워크 경량화 기술인 지식증류(knowledge distillation) 중 자가증류(self-distillation)를 초해상화 CNN 모델에 적용하여 성능을 평가, 분석한다. 실험 결과, 정량적 평가지표를 통하여 자가증류를 통해서도 성능이 우수한 경량화된 초해상화 모델을 얻을 수 있음을 확인하였다.

• Journal of Korea Multimedia Society
• /
• v.24 no.12
• /
• pp.1598-1605
• /
• 2021

Recently, convolutional neural networks (CNNs) have been widely used with excellent performance in various computer vision fields, including super-resolution (SR). However, CNN is computationally intensive and requires a lot of memory, making it difficult to apply to limited hardware resources such as mobile or Internet of Things devices. To solve these limitations, network lightening studies have been actively conducted to reduce the depth or size of pre-trained deep CNN models while maintaining their performance as much as possible. This paper aims to lighten the SR CNN model, SRGAN, using the knowledge distillation among network lightening technologies; thus, it proposes four techniques with different methods of transferring the knowledge of the teacher network to the student network and presents experiments to compare and analyze the performance of each technique. In our experimental results, it was confirmed through quantitative and qualitative evaluation indicators that student networks with knowledge transfer performed better than those without knowledge transfer, and among the four knowledge transfer techniques, the technique of conducting adversarial learning after transferring knowledge from the teacher generator to the student generator showed the best performance.

• Journal of Broadcast Engineering
• /
• v.25 no.2
• /
• pp.242-251
• /
• 2020

Recently, many deep convolutional neural networks for image super-resolution have been studied. Existing deep learning-based super-resolution algorithms are architecture that up-samples the resolution at the end of the network. The post-upsampling architecture has an inefficient structure at large scaling factor result of predicting a lot of information for mapping from low-resolution to high-resolution at once. In this paper, we propose a single image super-resolution using Channel Attention Residual Dense Block based on an iterative up-down sampling architecture. The proposed algorithm efficiently predicts the mapping relationship between low-resolution and high-resolution, and shows up to 0.14dB performance improvement and enhanced subjective image quality compared to the existing algorithm at large scaling factor result.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2017.11a
• /
• pp.781-784
• /
• 2017

Face recognition systems are characterized by low invasiveness of acquisition, and increasingly better reliability. Such systems may not be applied effectively, when the images are in low resolution (LR) as in the case that photos are taken from long distances, typically public surveillance. In theory, the high resolution (HR) image reconstructed from an LR face image, applying a super resolution (SR) method, can be used for face recognition. However, existing face SR algorithms may not give satisfactory results in face recognition. This article investigates the very low resolution face recognition problem and introduces a partial differential equation (PDE)-based SR method for a face recognition system of convolutional neural network (CNN).

• Journal of Broadcast Engineering
• /
• v.26 no.4
• /
• pp.429-440
• /
• 2021

With the recent development of deep convolutional neural network learning, deep learning techniques applied to single image super-resolution are showing good results. One of the existing deep learning-based super-resolution techniques is RDN(Residual Dense Network), in which the initial feature information is transmitted to the last layer using residual dense blocks, and subsequent layers are restored using input information of previous layers. However, if all hierarchical features are connected and learned and a large number of residual dense blocks are stacked, despite good performance, a large number of parameters and huge computational load are needed, so it takes a lot of time to learn a network and a slow processing speed, and it is not applicable to a mobile system. In this paper, we use the residual dense structure, which is a continuous memory structure that reuses previous information, and the residual dense channel attention block using the channel attention method that determines the importance according to the feature map of the image. We propose a method that can increase the depth to obtain a large receptive field and maintain a concise model at the same time. As a result of the experiment, the proposed network obtained PSNR as low as 0.205dB on average at 4× magnification compared to RDN, but about 1.8 times faster processing speed, about 10 times less number of parameters and about 1.74 times less computation.

• Geophysics and Geophysical Exploration
• /
• v.25 no.3
• /
• pp.140-161
• /
• 2022

Seismic data with missing traces are often obtained regularly or irregularly due to environmental and economic constraints in their acquisition. Accordingly, seismic data interpolation is an essential step in seismic data processing. Recently, research activity on machine learning-based seismic data interpolation has been flourishing. In particular, convolutional neural network (CNN) and generative adversarial network (GAN), which are widely used algorithms for super-resolution problem solving in the image processing field, are also used for seismic data interpolation. In this study, CNN-based algorithm, U-Net and GAN-based algorithm, and conditional Wasserstein GAN (cWGAN) were used as seismic data interpolation methods. The results and performances of the methods were evaluated thoroughly to find an optimal interpolation method, which reconstructs with high accuracy missing seismic data. The work process for model training and performance evaluation was divided into two cases (i.e., Cases I and II). In Case I, we trained the model using only the regularly sampled data with 50% missing traces. We evaluated the model performance by applying the trained model to a total of six different test datasets, which consisted of a combination of regular, irregular, and sampling ratios. In Case II, six different models were generated using the training datasets sampled in the same way as the six test datasets. The models were applied to the same test datasets used in Case I to compare the results. We found that cWGAN showed better prediction performance than U-Net with higher PSNR and SSIM. However, cWGAN generated additional noise to the prediction results; thus, an ensemble technique was performed to remove the noise and improve the accuracy. The cWGAN ensemble model removed successfully the noise and showed improved PSNR and SSIM compared with existing individual models.

• Korean Journal of Remote Sensing
• /
• v.38 no.2
• /
• pp.167-177
• /
• 2022

In order to obtain satellite image products using the image transmitted to the ground station after capturing the satellite images, many image pre/post-processing steps are involved. During the pre/post-processing, when converting from level 1R images to level 1G images, geometric correction is essential. An interpolation method necessary for geometric correction is inevitably used, and the quality of the level 1G images is determined according to the accuracy of the interpolation method. Also, it is crucial to speed up the interpolation algorithm by the level processor. In this paper, we proposed a lightweight CNN-based interpolation method required for geometric correction when converting from level 1R to level 1G. The proposed method doubles the resolution of satellite images and constructs a deep learning network with a lightweight deep convolutional neural network for fast processing speed. In addition, a feature map fusion method capable of improving the image quality of multispectral (MS) bands using panchromatic (PAN) band information was proposed. The images obtained through the proposed interpolation method improved by about 0.4 dB for the PAN image and about 4.9 dB for the MS image in the quantitative peak signal-to-noise ratio (PSNR) index compared to the existing deep learning-based interpolation methods. In addition, it was confirmed that the time required to acquire an image that is twice the resolution of the 36,500×36,500 input image based on the PAN image size is improved by about 1.6 times compared to the existing deep learning-based interpolation method.