• ETRI Journal
• /
• v.41 no.4
• /
• pp.415-425
• /
• 2019

Generative adversarial networks (GAN) have been successfully used in many image restoration tasks, including image denoising, super-resolution, and compression artifact reduction. By fully exploiting its characteristics, state-of-the-art image restoration techniques can be used to generate images with photorealistic details. However, there are many applications that require faithful rather than visually appealing image reconstruction, such as medical imaging, surveillance, and video coding. We found that previous GAN-training methods that used a loss function in the form of a weighted sum of fidelity and adversarial loss fails to reduce fidelity loss. This results in non-negligible degradation of the objective image quality, including peak signal-to-noise ratio. Our approach is to alternate between fidelity and adversarial loss in a way that the minimization of adversarial loss does not deteriorate the fidelity. Experimental results on compression-artifact reduction and super-resolution tasks show that the proposed method can perform faithful and photorealistic image restoration.

• Journal of Korea Multimedia Society
• /
• v.21 no.9
• /
• pp.1027-1034
• /
• 2018

CCTV has been known to possess high level of objectivity and utility. Hence, the government has recently focused on replacing low quality CCTV with higher quality ones or even by adding high resolution CCTV. However, converting all existing low-quality CCTV to high quality can be extremely costly. Furthermore, low quality videos prior to CCTV replacement are likely to be of poor quality and thus not utilized correctly. In order to solve these problems, this paper proposes a method to improve videos quality of images using SRGAN(Super Resolution Generative Advisory Networks). Through experiments, we have proven that it is possible to improve low quality CCTV videos clearly. For this experiment, a total of 4 types of CCTV videos were used and 10,000 images were sampled from each type. Those images could then be used for machine learning. The fact that the pre-process for machine learning has been done manually and the long time that required for machine learning seems to be complementary.

• Proceedings of the KOSOMBE Conference
• /
• v.1995 no.05
• /
• pp.164-166
• /
• 1995

The objectives of study are to develop a peripheral device on the endoscopic surgery system. These systems are consist of the following units. They are a color monitor of high resolution, light source, computer system and endoscopic camera with a C-mount head, irrigator, color video printer, Super VHS recorder and a system rack. The color monitor is a NTSC monitor for monitoring the image projected of the surgical section. The lightsource is necessary to irradiate the interior of a body via an optic fiber, The light projector will adapt the brightness in accordance with changing distance from the object. A miniature camera using a color CCD chip and computer system is used to capture and control an image of the surgical section[1]. The video printer is a 300 DPI resolution using thermal sublimation methods, which is developed by Samsung Electronics Co., Ltd. The specification of the endoscopic data management system is consist of storage of a captured image and pathological database of patients [2-4].

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2009.01a
• /
• pp.184-188
• /
• 2009

To interactively share High Definition (HD)-quality visualization over emerging ultra-high-speed network infrastructure, several lossless and low-delay real-time media (i.e., uncompressed HD video and audio) transport systems are being designed and prototyped. However, most of them still rely on expensive hardware components. As an effort to reduce the building cost of system, in this paper, we propose the integration of both transmitter and receiver machines into a single bi-directional transport system. After detailed bottleneck analysis and subsequent refinements of embedded software components, the proposed integration can provide Real-time Transport Protocol (RTP)-based bi-directional transport of uncompressed HD video and audio from a single machine. We also explain how to interface the Gbps-bandwidth display output of uncompressed HD media system to the networked tiled display of 10240 $\times$ 3200 super-high-resolution. Finally, to verify the feasibility of proposed integration, several prototype systems are built and evaluated by operating them in several different experiment scenarios.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2017.11a
• /
• pp.110-112
• /
• 2017

본 논문에서는 최신의 비디오 코덱 표준인 HEVC(High Efficiency Video Coding)의 화면 내 예측 부호화의 성능 향상을 위하여 SRCNN(Super Resolution Convolutional Neural Networks)을 이용하는 방법을 제안한다. SRCNN 은 비교적 최신 기술인 CNN(Convolutional Neural Network)을 사용하여 이미지를 추가적인 데이터 없이 보간 하여 해상도를 증가시키는 기술이다. HEVC 에서는 화면 내 예측의 잔차신호를 부호화 하기 위해 많은 비트를 소모하는데, 본 논문에서는 이 잔차신호들의 해상도를 낮추어 부호화 되는 비트를 줄이며, 복호화기에서 SRCNN 을 이용하여 원래의 해상도로 복원을 수행하여 압축성능을 향상 시키는 방법에 대하여 제안한다. 제안하는 기술은 HM 16.6 에 구현하였으며, CNN 트레이닝에 Caffe 라이브러리를 사용하였다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.06a
• /
• pp.1226-1228
• /
• 2022

FHD 이상을 넘어선 UHD급의 고해상도 동영상 콘텐츠의 수요 및 공급이 증가함에 따라 전반적인 산업 영역에서 네트워크 자원을 효율적으로 이용하여 동영상 콘텐츠를 제공하는 데에 관심을 두게 되었다. 기존 방법을 통한 bi-cubic, bi-linear interpolation 등의 방법은 딥 러닝 기반의 모델에 비교적 인풋 이미지의 특징을 잘 잡아내지 못하는 결과를 나타내었다. 딥 러닝 기반의 초 해상화 기술의 경우 기존 방법과 비교 시 연산을 위해 더 많은 자원을 필요로 하므로, 이러한 사용 조건에 따라 본 논문은 초 해상화가 가능한 딥 러닝 모델을 경량화 기법을 사용하여 기존에 사용된 모델보다 비교적 적은 자원을 효율적으로 사용할 수 있도록 연구 개발하는 데 목적을 두었다. 연구방법으로는 structure pruning을 이용하여 모델 자체의 구조를 경량화 하였고, 학습을 진행해야 하는 파라미터를 줄여 하드웨어 자원을 줄이는 연구를 진행했다. 또한, Residual Network의 개수를 줄여가며 PSNR, LPIPS, tOF등의 결과를 비교했다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.234-236
• /
• 2020

Content-adaptive training and transmission of the model parameters of neural networks can boost up the SR performance with higher restoration fidelity. In this case, efficient transmission of neural network parameters are essentially needed. Thus, we propose a novel method of compressing the network model parameters based on the training of network model parameters in the sense that the residues of filter parameters and content loss are jointly minimized. So, the residues of filter parameters are only transmitted to receiver sides for different temporal portions of video under consideration. This is advantage for image restoration applications with receivers (user terminals) of low complexity. In this case, the user terminals are assumed to have a limited computation and storage resource.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.237-240
• /
• 2020

The fine-tuned neural network (NN) model for a whole temporal portion in a video does not always yield the best quality (e.g., PSNR) performance over all regions of each frame in the temporal period. For certain regions (usually homogeneous regions) in a frame for super-resolution (SR), even a simple bicubic interpolation method may yield better PSNR performance than the fine-tuned NN model. When there are multiple NN models available at the receivers where each NN model is trained for a group of images having a specific category of image characteristics, the performance of Quality enhancement can be improved by selectively applying an appropriate NN model for each image region according to its image characteristic category to which the NN model was dedicatedly trained. In this case, it is necessary to signal which NN model is applied for each region. This is very advantageous for image restoration and quality enhancement (IRQE) applications at user terminals with limited computing capabilities.

• International conference on construction engineering and project management
• /
• 2024.07a
• /
• pp.1282-1282
• /
• 2024

Given the widespread use of intelligent surveillance cameras at construction sites, recent studies have introduced vision-based deep learning approaches. These studies have focused on enhancing the performance of vision-based excavator activity recognition to automatically monitor productivity metrics such as activity time and work cycle. However, acquiring a large amount of training data, i.e., videos captured from actual construction sites, is necessary for developing a vision-based excavator activity recognition model. Yet, complexities of dynamic working environments and security concerns at construction sites pose limitations on obtaining such videos from various surveillance camera locations. Consequently, this leads to performance degradation in excavator activity recognition models, reducing the accuracy and efficiency of heavy equipment productivity analysis. To address these limitations, this study aimed to conduct sensitivity analysis of excavator activity recognition performance based on surveillance camera location, utilizing synthetic videos generated from a game-engine-based virtual environment (Unreal Engine). Various scenarios for surveillance camera placement were devised, considering horizontal distance (20m, 30m, and 50m), vertical height (3m, 6m, and 10m), and horizontal angle (0° for front view, 90° for side view, and 180° for backside view). Performance analysis employed a 3D ResNet-18 model with transfer learning, yielding approximately 90.6% accuracy. Main findings revealed that horizontal distance significantly impacted model performance. Overall accuracy decreased with increasing distance (76.8% for 20m, 60.6% for 30m, and 35.3% for 50m). Particularly, videos with a 20m horizontal distance (close distance) exhibited accuracy above 80% in most scenarios. Moreover, accuracy trends in scenarios varied with vertical height and horizontal angle. At 0° (front view), accuracy mostly decreased with increasing height, while accuracy increased at 90° (side view) with increasing height. In addition, limited feature extraction for excavator activity recognition was found at 180° (backside view) due to occlusion of the excavator's bucket and arm. Based on these results, future studies should focus on enhancing the performance of vision-based recognition models by determining optimal surveillance camera locations at construction sites, utilizing deep learning algorithms for video super resolution, and establishing large training datasets using synthetic videos generated from game-engine-based virtual environments.