• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.3
• /
• pp.109-114
• /
• 2021

GAN(Generative Adversarial Network) is an image generation model, which is composed of a generator network and a discriminator network, and generates an image similar to a real image. Since the image generated by the GAN should be similar to the actual image, a loss function is used to minimize the loss error of the generated image. However, there is a problem that the loss function of GAN degrades the quality of the image by making the learning to generate the image unstable. To solve this problem, this paper analyzes GAN-related studies and proposes an edge GAN(eGAN) using edge detection. As a result of the experiment, the eGAN model has improved performance over the existing GAN model.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2022.01a
• /
• pp.283-284
• /
• 2022

본 논문은 TTS를 사용해 사용자들에게 E-Book 및 뉴스를 보고 들을 수 있는 기능을 제공한다. 사용자 및 개발자가 직접 녹음한 TTS를 사용해 원하는 목소리, 배속과 같은 기능을 제공한다. 기존 TTS를 사용한 E-Book 사이트들은 광고가 많아 가독성의 문제와 유료 서비스인 반면에 본 논문에서 제안한 웹은 다양한 연령층의 사용자들이 사용하기 쉽게 메뉴의 간편화를 통해 다양한 E-Book 및 뉴스 기능을 제공함으로써 보다 직관적이고 쉽게 전자문서를 읽을 수 있도록 하였다.

• Journal of Broadcast Engineering
• /
• v.25 no.2
• /
• pp.143-150
• /
• 2020

Huge amounts of contents are being uploaded every day on various streaming platforms. Among those videos, game and sports videos account for a great portion. The broadcasting companies sometimes create and provide highlight videos. However, these tasks are time-consuming and costly. In this paper, we propose models that automatically predict highlights in games and sports matches. While most previous approaches use visual information exclusively, our models use both audio and visual information, and present a way to understand short term and long term flows of videos. We also describe models that combine GAN to find better highlight features. The proposed models are evaluated on e-sports and baseball videos.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.1
• /
• pp.9-18
• /
• 2021

Although Generative Adversarial Networks (GANs) have gained great popularity in computer vision and related fields, generating audio signals independently has yet to be presented. Unlike images, an audio signal is a sampled signal consisting of discrete samples, so it is not easy to learn the signals using CNN architectures, which is widely used in image generation tasks. In order to overcome this difficulty, GAN researchers proposed a strategy of applying time-frequency representations of audio to existing image-generating GANs. Following this strategy, we propose an improved method for increasing the fidelity of synthesized audio signals generated by using GANs. Our method is demonstrated on a public speech dataset, and evaluated by Fréchet Inception Distance (FID). When employing our method, the FID showed 10.504, but 11.973 as for the existing state of the art method (lower FID indicates better fidelity).

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.9
• /
• pp.1271-1274
• /
• 2021

RF fingerprinting based on deep learning (DL) has gained interests as a means to improve the security of near field communication (NFC) by allowing identification of NFC tags based on unique physical characteristics. To achieve high accuracy in the identification of NFC tags, it is crucial to utilize a large number of training data, however it is hard to collect such dataset in practice. In this study, we have provided new methodology to generate RF waveform from NFC tags, i.e., data augmentation, based on a conditional generative adversarial network (CGAN). By using the RF waveform of NFC tags which is collected from the testbed with software defined radio (SDR), we have confirmed that the realistic RF waveform can be generated through our proposed scheme.

• Journal of the Institute of Convergence Signal Processing
• /
• v.23 no.2
• /
• pp.97-106
• /
• 2022

In this study, H&E staining is necessary to distinguish cells. However, dyeing directly requires a lot of money and time. The purpose is to convert the phase image of unstained cells to the amplitude image of stained cells. Image data taken with FPM was created with Phase image and Amplitude image using Matlab's parameters. Through normalization, a visually identifiable image was obtained. Through normalization, a visually distinguishable image was obtained. Using the GAN algorithm, a Fake Amplitude image similar to the Real Amplitude image was created based on the Phase image, and cells were distinguished by objectification using MASK R-CNN with the Fake Amplitude image As a result of the study, D loss max is 3.3e-1, min is 6.8e-2, G loss max is 6.9e-2, min is 2.9e-2, A loss max is 5.8e-1, min is 1.2e-1, Mask R-CNN max is 1.9e0, and min is 3.2e-1.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.7
• /
• pp.2304-2320
• /
• 2021

Intelligently detecting anomalies in health sensor data streams (e.g., Electrocardiogram, ECG) can improve the development of E-health industry. The physiological signals of patients are collected through sensors. Timely diagnosis and treatment save medical resources, promote physical health, and reduce complications. However, it is difficult to automatically classify the ECG data, as the features of ECGs are difficult to extract. And the volume of labeled ECG data is limited, which affects the classification performance. In this paper, we propose a Generative Adversarial Network (GAN)-based deep learning framework (called CAB) for heart arrhythmia classification. CAB focuses on improving the detection accuracy based on a small number of labeled samples. It is trained based on the class-imbalance ECG data. Augmenting ECG data by a GAN model eliminates the impact of data scarcity. After data augmentation, CAB classifies the ECG data by using a Bidirectional Long Short Term Memory Recurrent Neural Network (Bi-LSTM). Experiment results show a better performance of CAB compared with state-of-the-art methods. The overall classification accuracy of CAB is 99.71%. The F1-scores of classifying Normal beats (N), Supraventricular ectopic beats (S), Ventricular ectopic beats (V), Fusion beats (F) and Unclassifiable beats (Q) heartbeats are 99.86%, 97.66%, 99.05%, 98.57% and 99.88%, respectively. Unclassifiable beats (Q) heartbeats are 99.86%, 97.66%, 99.05%, 98.57% and 99.88%, respectively.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.7
• /
• pp.2547-2567
• /
• 2021

It has been widely acknowledged that occlusion impairments adversely distress many face recognition algorithms' performance. Therefore, it is crucial to solving the problem of face image occlusion in face recognition. To solve the image occlusion problem in face recognition, this paper aims to automatically de-occlude the human face majority or discriminative regions to improve face recognition performance. To achieve this, we decompose the generative process into two key stages and employ a separate generative adversarial network (GAN)-based network in both stages. The first stage generates an initial coarse face image without an occlusion mask. The second stage refines the result from the first stage by forcing it closer to real face images or ground truth. To increase the performance and minimize the artifacts in the generated result, a new refine loss (e.g., reconstruction loss, perceptual loss, and adversarial loss) is used to determine all differences between the generated de-occluded face image and ground truth. Furthermore, we build occluded face images and corresponding occlusion-free face images dataset. We trained our model on this new dataset and later tested it on real-world face images. The experiment results (qualitative and quantitative) and the comparative study confirm the robustness and effectiveness of the proposed work in removing challenging occlusion masks with various structures, sizes, shapes, types, and positions.

• Journal of Information Processing Systems
• /
• v.18 no.4
• /
• pp.575-586
• /
• 2022

Uneven light in real-world causes visual degradation for underexposed regions. For these regions, insufficient consideration during enhancement procedure will result in over-/under-exposure, loss of details and color distortion. Confronting such challenges, an unsupervised low-light image enhancement network is proposed in this paper based on the guidance of the unpaired low-/normal-light images. The key components in our network include super-resolution module (SRM), a GAN-based low-light image enhancement network (LLIEN), and denoising-scaling module (DSM). The SRM improves the resolution of the low-light input images before illumination enhancement. Such design philosophy improves the effectiveness of texture details preservation by operating in high-resolution space. Subsequently, local lightness attention module in LLIEN effectively distinguishes unevenly illuminated areas and puts emphasis on low-light areas, ensuring the spatial consistency of illumination for locally underexposed images. Then, multiple discriminators, i.e., global discriminator, local region discriminator, and color discriminator performs assessment from different perspectives to avoid over-/under-exposure and color distortion, which guides the network to generate images that in line with human aesthetic perception. Finally, the DSM performs noise removal and obtains high-quality enhanced images. Both qualitative and quantitative experiments demonstrate that our approach achieves favorable results, which indicates its superior capacity on illumination and texture details restoration.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.30 no.5
• /
• pp.945-956
• /
• 2020

The publication of data is essential in order to receive high quality services from many applications. However, if the original data is published as it is, there is a risk that sensitive information (political tendency, disease, ets.) may reveal. Therefore, many research have been proposed, not the original data but the synthetic data generating and publishing to privacy preserve. but, there is a risk of privacy leakage still even if simply generate and publish the synthetic data by various attacks (linkage attack, inference attack, etc.). In this paper, we propose a synthetic data generation algorithm in which privacy preserved by applying differential privacy the latest privacy protection technique to GAN, which is drawing attention as a synthetic data generative model in order to prevent the leakage of such sensitive information. The generative model used CGAN for efficient learning of labeled data, and applied Rényi differential privacy, which is relaxation of differential privacy, considering the utility aspects of the data. And validation of the utility of the generated data is conducted and compared through various classifiers.