• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.11a
• /
• pp.76-79
• /
• 2019

합성곱 신경망의 성능이 증가하면서 다양한 영상 처리 문제를 해결하기 위해 합성곱 신경망을 적용한 시도들이 증가하고 있다. 고해상도 복원 문제도 그 중 하나였으며, 보다 높은 성능을 얻기 위해 주로 신경망의 깊이를 깊게 하는 시도들이 있었다. 본 논문에서는 고해상도 복원 작업을 위한 합성곱 신경망의 성능 향상을 위해 깊이를 증가시키는 접근법이 아닌 수용영역을 확장시키는 접근법을 시도하였다. 논문에서 제시한 모델은 신경망 내부에 두 개의 브랜치를 두어, 하나의 브랜치는 Dilated Convolution 을 이용해 수용영역을 확장하는데 사용되며, 다른 하나는 이 브랜치를 통해 나온 feature 를 가공하는데 사용된다. 기본 모델은 EDSR 을 사용하였으며, 최종적으로 4.79M 의 파라미터로 평균 32.46dB 의 PSNR 을 보여주었다. 하지만 모델의 구조가 복잡하여 깊이를 늘이는 접근법을 적용하기 어렵다는 한계점이 있다.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2021.01a
• /
• pp.65-68
• /
• 2021

본 연구에서는 Dilated U-Net 기반의 이미지 복원기법을 통해 콘크리트 균열 추출 성능 개선 방안을 제안한다. 콘크리트 균열은 구조물의 미관상의 문제뿐 아니라 추후 큰 안전사고의 원인이 될 수 있어 초기대응이 중요하다. 현재는 점검자가 직접 육안으로 검사하는 외관 검사법이 주로 사용되고 있지만, 이는 정확성 및 비용, 시간, 그리고 안전성 면에서 한계를 갖고 있다. 이에 콘크리트 구조물 표면에 대해 획득한 영상 처리 기법을 사용한 검사 방식 도입의 관심이 늘어나고 있다. 또한, 딥러닝 기술의 발달로 딥러닝을 적용한 영상처리의 연구 역시 활발하게 진행되고 있다. 본 연구는 콘크리트 균열 추개선출 성능 개선을 위해 Dilated U-Net 기반의 이미지 복원기법을 적용하는 방안을 제안하였고 성능 검증 결과, 기존 U-Net 기반의 정확도가 98.78%, 조화평균 82.67%였던 것에 비해 정확도 99.199%, 조화평균 88.722%로 성능이 되었음을 확인하였다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.139-143
• /
• 2021

본 논문에서는 꾸준히 연구되어 오던 이미지 복원 문제에서 초해상화와 인페인팅이라는 복합적 이미지 복원을 동시에 처리하는 해결 방법을 제안한다. 초해상화는 국지적 픽셀 정보를 이용하여 고해상도의 영상을 복원하고, 인페인팅은 이미지 전체 정보를 활용하여 영상 내 비어 있는 영역을 생성해야 하므로, 이러한 두 가지 영상 복원 기법을 동시에 수행하는 것은 상당히 어려운 문제이다. 그렇기에 인페인팅과 초해상화는 이미지 복원에서 널리 활용되는 기술인 만큼 동시에 해결할 수 있는 기법에 대한 수요는 있음에도 지금까지 거의 연구되지 않았다. 본 논문은 초해상화 및 인페인팅 합동 처리에 있어 복합적인 정보를 모두 다뤄야하는 네트워크가 서로의 성능을 저하시키지 않도록 개략적 복원 네트워크 (Coarse network), 디테일 복원 네트워크 (Refinement network), 초해상화 네트워크 (SR network)로 분리하여 초해상화 및 인페인팅 합동 처리를 수행하며, 각 단계마다 결과 영상을 얻어 스케일 별 정답 영상과 손실함수를 계산하여 복합적인 성능을 올릴 수 있는 방법을 제시한다. 또한 순차적 단일 모델에 비하여 인페인팅과 초해상화를 합동 학습하는 제안 모델이 개선된 화질의 결과 영상을 획득할 수 있다는 것을 실험적으로 보인다.

• Journal of IKEEE
• /
• v.20 no.2
• /
• pp.152-162
• /
• 2016

Electrical resistivity tomography (ERT) is a technique to reconstruct the internal resistivity distribution using the measured voltages on the surface electrodes. ERT inverse problem suffers from ill-posedness nature, so regularization methods are used to mitigate ill-posedness. The reconstruction performance varies depending on the type of regularization method. In this paper, an interacting dual-mode regularization method is proposed with two different regularization methods, L1-norm regularization and total variation (TV) regularization, to achieve robust reconstruction performance. The interacting dual-mode regularization method selects the suitable regularization method and combines the regularization methods based on computed mode probabilities depending on the actual conditions. The proposed method is tested with numerical simulations and the results demonstrate an improved reconstruction performance.

• Journal of Broadcast Engineering
• /
• v.25 no.6
• /
• pp.898-910
• /
• 2020

Single image-based super-resolution has been studied for a long time in computer vision because of various applications. Various deep learning-based super-resolution algorithms are introduced recently to improve the performance by reducing side effects like blurring and staircase effects. Most deep learning-based approaches have focused on how to implement the network architecture, loss function, and training strategy to improve performance. Meanwhile, Several approaches using Attention Module, which emphasizes the extracted features, are introduced to enhance the performance of the network without any additional layer. Attention module emphasizes or scales the feature map for the purpose of the network from various perspectives. In this paper, we propose the various channel attention and spatial attention in single image-based super-resolution and analyze the results and performance according to the architecture of the attention module. Also, we explore that designing multi-attention module to emphasize features efficiently from various perspectives.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.52 no.3
• /
• pp.154-161
• /
• 2015

A novel multiframe super-resolution (SR) algorithm is presented to overcome the limitation of existing single-image SR algorithms using motion information from adjacent frames in a video. The proposed SR algorithm consists of three steps: i) definition of a local region using interframe motion vectors, ii) multiscale patch generation and adaptive selection of multiple optimum patches, and iii) combination of optimum patches for super-resolution. The proposed algorithm increases the accuracy of patch selection using motion information and multiscale patches. Experimental results show that the proposed algorithm performs better than existing patch-based SR algorithms in the sense of both subjective and objective measures including the peak signal-to-noise ratio (PSNR) and structural similarity measure (SSIM).

• Journal of KIISE:Software and Applications
• /
• v.34 no.5
• /
• pp.476-486
• /
• 2007

This paper proposes a resolution enhancement method which can reconstruct high-resolution facial images from single-frame, low-resolution facial images. The proposed method is derived from example-based reconstruction methods and the morphable face model. In order to improve the performance of the example-based reconstruction, we propose the region-based reconstruction method which can maintain the characteristics of local facial regions. Also, in order to use the capability of the morphable face model to face resolution enhancement problems, we define the extended morphable face model in which an extended face is composed of a low-resolution face, its interpolated high-resolution face, and the high-resolution equivalent, and then an extended face is separated by an extended shape vector and an extended texture vector. The encouraging results show that the proposed methods can be used to improve the performance of face recognition systems, particularly to enhance the resolution of facial images captured from visual surveillance systems.

• Journal of the Korea Computer Graphics Society
• /
• v.27 no.3
• /
• pp.65-74
• /
• 2021

3D reconstruction means that reconstructing the 3D shape of the object in an image and a video. We proposed a progressive occupancy network architecture that can recover not only the overall shape of the object but also the local details. Unlike the original occupancy network, which uses a feature vector embedding information of the whole image, we extract and utilize the different levels of image features depending on the receptive field size. We also propose a novel network architecture that applies the image features sequentially to the decoder blocks in the decoder and improves the quality of the reconstructed 3D shape progressively. In addition, we design a novel decoder block structure that combines the different levels of image features properly and uses them for updating the input point feature. We trained our progressive occupancy network with ShapeNet. We compare its representation power with two prior methods, including prior occupancy network(ONet) and the recent work(DISN) that used different levels of image features like ours. From the perspective of evaluation metrics, our network shows better performance than ONet for all the metrics, and it achieved a little better or a compatible score with DISN. For visualization results, we found that our method successfully reconstructs the local details that ONet misses. Also, compare with DISN that fails to reconstruct the thin parts or occluded parts of the object, our progressive occupancy network successfully catches the parts. These results validate the usefulness of the proposed network architecture.

• Journal of Broadcast Engineering
• /
• v.27 no.3
• /
• pp.351-360
• /
• 2022

Image blurring is a phenomenon that occurs due to factors such as movement of a subject and shaking of a camera. Recently, the research for image deblurring has been actively conducted based on convolution neural networks. In particular, the method of guiding the restoration process via the difference between blur and sharp images has shown the promising performance. This paper proposes a novel method for improving the deblurring performance based on the distortion information. To this end, the transformer-based neural network module is designed to guide the restoration process. The proposed method efficiently reflects the distorted region, which is predicted through the global inference during the deblurring process. We demonstrate the efficiency and robustness of the proposed module based on experimental results with various deblurring architectures and benchmark datasets.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.18 no.5
• /
• pp.895-900
• /
• 2023

We study low-complexity graph sampling which selects a subset of nodes from graph nodes so as to reconstruct the original signal from the sampled one. To achieve complexity reduction, we propose a graph sampling algorithm with thresholding which selects a node with a cost lower than a given threshold at each step without fully searching all of the remaining nodes to find one with the minimum cost. Since it is important to find the threshold as close to a minimum cost as possible to avoid degradation of the reconstruction performance, we present a mathematical expression to compute the threshold at each step. We investigate the performance of the different sampling methods for various graphs, showing that the proposed algorithm runs 1.3 times faster than the previous method while maintaining the reconstruction performance.