• Journal of the Institute of Convergence Signal Processing
• /
• v.23 no.4
• /
• pp.222-227
• /
• 2022

In this paper, we implemented a Korean text generation and classification model based on a deep learning algorithm that can be applied to various industries. It consists of two implemented GAN-based Korean handwriting generation models and CNN-based Korean handwriting classification models. The GAN model consists of a generator model for generating fake Korean handwriting data and a discriminator model for discriminating fake handwritten data. In the case of the CNN model, the model was trained using the 'PHD08' dataset, and the learning result was 92.45. It was confirmed that Korean handwriting was classified with % accuracy. As a result of evaluating the performance of the classification model by integrating the Korean cursive data generated through the implemented GAN model and the training dataset of the existing CNN model, it was confirmed that the classification performance was 96.86%, which was superior to the existing classification performance.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2024.05a
• /
• pp.69-72
• /
• 2024

모델 역추론 공격은 공격 대상 네트워크를 훈련하기 위해 사용되는 훈련 데이터셋 중 개인 데이터셋을 공개 데이터셋을 사용하여 개인 훈련 데이터셋을 복원하는 것이다. 모델 역추론 방법 중 적대적 생성 신경망을 사용하여 모델 역추론 공격을 하는 과거의 논문들은 딥러닝 모델 전체의 역추론에만 초점을 맞추기 때문에, 이를 통해 얻은 원본 이미지의 개인 데이터 정보는 제한적이다. 따라서, 본 연구는 대상 모델의 중간 출력을 사용하여 개인 데이터에 대한 더 품질 높은 정보를 얻는데 초점을 맞춘다. 본 논문에서는 적대적 생성 신경망 모델이 원본 이미지를 생성하기 위해 사용되는 계층별 역추론 공격 방법을 소개한다. MNIST 데이터셋으로 훈련된 적대적 생성 신경망 모델을 사용하여, 원본 이미지가 대상 모델의 계층을 통과하면서 얻은 중간 계층의 출력 데이터를 기반으로 원본 이미지를 재구성하고자 한다. GMI 의 공격 방식을 참고하여 공격 모델의 손실 함수를 구성한다. 손실 함수는 사전 손실 및 정체성 손실항을 포함하며, 역전파를 통해서 원본 이미지와 가장 유사하게 복원할 수 있는 표현 벡터 Z 를 찾는다. 원본 이미지와 공격 이미지 사이의 유사성을 분류 라벨의 정확도, SSIM, PSNR 값이라는 세 가지 지표를 사용하여 평가한다. 공격이 이루어지는 계층에서 복원한 이미지와 원본 이미지를 세 가지 지표를 가지고 평가한다. 실험 결과, 공격 이미지가 원본 이미지의 대상 분류 라벨을 정확하게 가지며 원본 이미지의 필체를 유사하게 복원하였음을 보여준다. 평가 지표 또한 원본 이미지와 유사하다는 것을 나타낸다.

• Journal of Broadcast Engineering
• /
• v.13 no.4
• /
• pp.544-553
• /
• 2008

With the development of video editing technology, there are growing uses of overlay text inserted into video contents to provide viewers with better visual understanding. Since the content of the scene or the editor's intention can be well represented by using inserted text, it is useful for video information retrieval and indexing. Most of the previous approaches are based on low-level features, such as edge, color, and texture information. However, existing methods experience difficulties in handling texts with various contrasts or inserted in a complex background. In this paper, we propose a novel framework to localize the overlay text in a video scene. Based on our observation that there exist transient colors between inserted text and its adjacent background a transition map is generated. Then candidate regions are extracted by using the transition map and overlay text is finally determined based on the density of state in each candidate. The proposed method is robust to color, size, position, style, and contrast of overlay text. It is also language free. Text region update between frames is also exploited to reduce the processing time. Experiments are performed on diverse videos to confirm the efficiency of the proposed method.

• Smart Media Journal
• /
• v.12 no.1
• /
• pp.24-31
• /
• 2023

Meta-learning is metacognition that instantly distinguishes between knowing and unknown. It is a learning method that adapts and solves new problems by self-learning with a small amount of data.A few-shot learning method is a type of meta-learning method that accurately predicts query data even with a very small support set. In this study, we propose a method to solve the limitations of the prototype created with the mean-point vector of each class. For this purpose, we use the few-shot learning method that created the prototype used in the few-shot learning method as the median prototype. For quantitative evaluation, a handwriting recognition dataset and mini-Imagenet dataset were used and compared with the existing method. Through the experimental results, it was confirmed that the performance was improved compared to the existing method.

• Journal of KIISE:Computing Practices and Letters
• /
• v.15 no.5
• /
• pp.365-369
• /
• 2009

Overlapping windows are generally used for queries to process continuous data streams. Nevertheless, existing approaches discussed join algorithms only for basic types of windows such as tumbling windows and tuple-driven windows. In this paper, we propose an efficient join algorithm for overlapping windows, which are considered as a more general type of windows. The proposed algorithm is based on an incremental window join. It focuses on producing join results continuously when the memory overflow frequently occurs. It consists of (1) a method to use both of the incremental and full joins selectively, (2) a victim selection algorithm to minimize latency of join processing and (3) an idle time professing algorithm. We show through our experiments that the selective use of incremental and full joins provides better performance than using one of them only.