• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2021.06a
• /
• pp.345-347
• /
• 2021

단일 이미지 초해상도는 딥러닝의 발전과 함께 놀라운 성능 향상이 이루어 졌다. 이러한 딥러닝 모델은 매우 많은 파라미터를 갖고 있어 많은 연산량과 메모리를 필요로 한다. 하지만 사용할 수 있는 리소스는 한정되어 있기 때문에 네트워크를 경량화 시키려는 연구도 지속되어 왔다. 본 논문에서는 Stochastic Weight Averaging (SWA) 알고리즘을 이용하여 상대적으로 적은 양의 메모리와 연산을 추가해 이미지 초해상도 모델의 성능을 높이고 안정적인 학습을 달성하였다. SWA 알고리즘을 적용한 모델은 그렇지 않은 모델에 비해 테스트셋에서 최대 0.13dB 의 성능 향상을 보였다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.656-659
• /
• 2020

초해상화 딥러닝 기법은 학습 시 수렴하기까지 최소 수백 번의 에폭을 필요로 하며 오랜 시간이 걸린다. 최근, 영상 인식용 딥러닝 모델에서는 학습 수렴 속도를 향상시키기 위해 픽셀, 채널간 불필요한 상호연관 정보를 제거하는 Deconvolution 기술이 제안되었다. 본 논문에서는 최초로 Deconvolution 기술을 초해상화 딥러닝 방법에 적용하여 학습 수렴 속도 증가를 시도했다. 영상 인식 딥러닝 기법과 다르게 초해상화 딥러닝 기법은 이미지 특성 추출 부분과 이미지 복원 부분의 정보를 보존하는 것이 중요하기 때문에, EDSR을 Baseline 모델로 사용하여 양쪽 끝의 레이어는 기존의 Convolution 연산을 그대로 유지하고, 중간 레이어의 ResBlock 내의 Convolution 연산만 Deconvolution 연산으로 바꿔서 구성하였다. 초해상화 벤치마크 데이터셋을 사용한 실험 결과, 수렴속도가 빨라지지 않는 결과를 도출했다. 본 논문에서는 Deconvolution 기술이 Baseline 모델의 성능을 개선하지 못하는 이유를 초해상화 분야에서 기본적으로 적용되는 Residual Learning 기법 때문으로 분석했다.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.129-132
• /
• 2022

Super-resolution is the process of converting a low-quality image into a high-quality image. This study was conducted using ESPCN. In a super-resolution deep neural network, different quality images can be output even when receiving the same input data according to the activation function that determines the weight when passing through each node. Therefore, the purpose of this study is to find the most suitable activation function for super-resolution by applying the activation functions ReLU, ELU, and Swish and compare the quality of the output image for the same input images. The CelebaA Dataset was used as the dataset. Images were cut into a square during the pre-processing process then the image quality was lowered. The degraded image was used as the input image and the original image was used for evaluation. As a result, ELU and swish took a long time to train compared to ReLU, which is mainly used for machine learning but showed better performance.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.11
• /
• pp.1477-1485
• /
• 2021

Recently, super-resolution has been intensively studied only on upscaling models with integer magnification. However, the need to expand arbitrary magnification is emerging in representative application fields of actual super-resolution, such as object recognition and display image quality improvement. In this paper, we propose a model that can support arbitrary magnification by using the weights of the existing integer magnification model. This model converts super-resolution results into the DCT spectral domain to expand the space for arbitrary magnification. To reduce the loss of high-frequency information in the image caused by the expansion by the DCT spectral domain, we propose a high-frequency attention network for arbitrary magnification so that this model can properly restore high-frequency spectral information. To recover high-frequency information properly, the proposed network utilizes channel attention layers. This layer can learn correlations between RGB channels, and it can deepen the model through residual structures.

• Journal of the Korea Computer Graphics Society
• /
• v.29 no.5
• /
• pp.31-40
• /
• 2023

As display resolution increases, many apps also tend to include high-resolution texture maps. Recent advancements in deep-learning-based image super-resolution techniques make it possible to automate high-resolution texture generation. However, there is still a lack of comprehensive analysis of the application of these techniques to texture maps. In this paper, we selected three recent super-resolution techniques, namely BSRGAN, Real-ESRGAN, and SwinIR (classical and real-world image SR), and applied them to upscale texture maps. We then conducted a quantitative and qualitative analysis of the experimental results. The findings revealed various artifacts after upscaling, which indicates that there are still limitations in directly applying super-resolution techniques to texture-map upscaling.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.337-339
• /
• 2020

오디오 초 해상도 기술은 저 해상도의 오디오 신호를 이용하여 고 해상도의 오디오를 복원 또는 생성해 내는 기술이다. 본 기술 분야는 기존에 주파수 대역 확장, 인공 대역 확장 기술 등으로 연구되었으나, 최근 딥러닝 기술의 발전, 이미지 초 해상도 기술 연구 등에 힘입어 오디오 초 해상도 기술 이라는 이름으로 주로 연구되고 있다. 본 논문에서는 이러한 오디오 초 해상도 기술에 연구 동향에 대하여 설명하고, 기존의 논문 들에서 주로 다루고 있는 음성 데이터 베이스가 아닌 MedleyDB 음악 데이터 베이스를 활용하여 실험을 수행하였다. 실험은 4-폴드 교차 검증을 통해 수행되었으며, 실험 결과 제안하는 컨벌루션 신경망 구조 기반 오디오 초 해상도 기술은 입력 저해상도 오디오 대비 SNR 이 3.41 dB 향상됨을 확인하였다.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.4
• /
• pp.543-548
• /
• 2021

The higher the resolution of the image, the higher the satisfaction of the viewers of the image, and the super-resolution imaging has a considerable increase in research value among the fields of computer vision and image processing. In this study, the main features of low-resolution image LR are extracted mainly using deep learning super-resolution models. It learns and reconstructs the extracted features, and focuses on reconstruction-based algorithms that generate high-resolution image HR. In this paper, we investigate SRCNN and VDSR in a super-resolution algorithm model based on reconstruction. The structure and algorithm process of the SRCNN and VDSR model are briefly introduced, and the multi-channel and special form are also examined in the improved performance evaluation function, and understand the performance of each algorithm through experiments. In the experiment, an experiment was performed to compare the results of the SRCNN and VDSR models with the peak signal-to-noise ratio and image structure similarity, so that the results can be easily judged.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.11
• /
• pp.1486-1494
• /
• 2021

Character recognition is a technology required in various platforms, such as smart parking and text to speech, and many studies are being conducted to improve its performance through new attempts. However, with low-quality image used for character recognition, a difference in resolution of the training image and test image for character recognition occurs, resulting in poor accuracy. To solve this problem, this paper designed an end-to-end learning neural network that combines image super-resolution and character recognition so that the character recognition model performance is robust against various quality data, and implemented an alternative whole learning algorithm to learn the whole neural network. An alternative end-to-end learning and recognition performance test was conducted using the license plate image among various text images, and the effectiveness of the proposed algorithm was verified with the performance test.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.193-196
• /
• 2021

홀로그램(Hologram)은 3차원 물체에서 나오는 빛의 정보를 제어하는 기술이다. 현재는 컴퓨터 생성 홀로그램(CGH)으로 생성한 디지털 홀로그램에 관한 연구, 특히 물체에서 나오는 빛의 정보를 최대한 기록하고 재현하여 디지털 홀로그램의 해상도를 향상 시키려는 연구가 활발히 진행되고 있다. 이에 본 논문에서는 고해상도 홀로그램 영상을 얻기 위해 딥러닝 기반 초해상도(Super Resolution) 네트워크를 훈련 및 최적화하여, 저해상도 위상 홀로그램 영상으로부터 높은 화질의 홀로그램 영상을 재현하는 고해상도 위상 홀로그램 영상을 생성하는 것을 목표로 한다. 이때 위상 홀로그램 영상의 특성을 이용한 순환 손실 함수(Circular loss function)를 새롭게 제안하며, 기존의 이미지 초해상도 신경망 모델을 학습시킬 때 자주 사용하는 L1 손실 함수와 비교했을 때 약 0.13dB 정도의 성능 향상이 있었다.

• Journal of Broadcast Engineering
• /
• v.28 no.2
• /
• pp.178-184
• /
• 2023

Finding texts in general scene images and recognizing their contents is a very important task that can be used as a basis for robot vision, visual assistance, and so on. However, for the low-resolution text images, the degradations, such as noise or blur included in text images, are more noticeable, which leads to severe performance degradation of text recognition accuracy. In this paper, we propose a new Korean text image super-resolution based on a Transformer-based model, which generally shows higher performance than convolutional neural networks. In the experiments, we show that text recognition accuracy for Korean text images can be improved when our proposed text image super-resolution method is used. We also propose a new Korean text image dataset for training our model, which contains massive HR-LR Korean text image pairs.