• Proceedings of the Korean Society of Computer Information Conference
• /
• 2023.01a
• /
• pp.17-18
• /
• 2023

비지도형 도메인 적응(UDA)에 대한 최근 연구는 도메인 적응에 대한 설명 및 전이 가능한 특징을 풀어 내기 위해 적대적 학습에 의존한다. 그러나 기존 방법에는 대상 도메인의 클래스 인식(class-aware) 정보를 고려하지 않고는 잠재 공간의 구별 가능성을 완전히 보장할 수 없다는 것과 소스 및 대상 도메인의 샘플만으로는 잠재 공간에서 도메인 불변(domain- invariant) 특성을 추출하기에 부족하다는 두 가지 문제가 있다고 알려져 있다. 본 논문에서는 기존 알려진 UDA의 도메인 적응시 발생되는 문제를 해결하기 위해 Adversarial Discriminative Domain Adaptation(ADDA)에서 mixup을 활용해 신경망의 로버스트네스를 향상시키는 것을 확인하였다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2015.04a
• /
• pp.687-690
• /
• 2015

본 논문에서는 대칭 변환을 지원하는 윤곽선 이미지 매칭 문제를 다룬다. 이미지 매칭에서 이미지의 대칭 변환을 지원하는 것은 직관적이고 정확한 매칭을 위해 매우 중요한 요소이다. 그러나, 기존 이미지 매칭에서는 이미지 회전-불변만 고려하였을 뿐 대칭 변환은 고려하지 않았다. 본 논문에서는 기존 회전-불변 윤곽선 이미지 매칭과 함께 대칭 변환을 지원하는 새로운 윤곽선 이미지 매칭을 제안한다. 실험 결과, 제안하는 대칭 변환 매칭은 회전-불변만을 지원하는 기존 윤곽선 이미지 매칭에 비해 보다 정확하고 직관적인 결과를 도출하는 것으로 나타났다. 이 같은 결과는 대칭 변환 윤곽선 이미지 매칭이 이미지의 대칭 변환 문제를 시계열 도메인에서 해결한 우수한 해결책임을 의미한다.

• Journal of the Korea Institute of Military Science and Technology
• /
• v.24 no.4
• /
• pp.382-391
• /
• 2021

Unsupervised domain adaptation often gives impressive solutions to handle domain shift of data. Most of current approaches assume that unlabeled target data to train is abundant. This assumption is not always true in practices. To tackle this issue, we propose a general solution to solve the domain gap minimization problem without any target data. Our method consists of two regularization steps. The first step is a pixel regularization by arbitrary style transfer. Recently, some methods bring style transfer algorithms to domain adaptation and domain generalization process. They use style transfer algorithms to remove texture bias in source domain data. We also use style transfer algorithms for removing texture bias, but our method depends on neither domain adaptation nor domain generalization paradigm. The second regularization step is a feature regularization by feature alignment. Adding a feature alignment loss term to the model loss, the model learns domain invariant representation more efficiently. We evaluate our regularization methods from several experiments both on small dataset and large dataset. From the experiments, we show that our model can learn domain invariant representation as much as unsupervised domain adaptation methods.

• KIPS Transactions on Software and Data Engineering
• /
• v.4 no.10
• /
• pp.431-438
• /
• 2015

In this paper we address the symmetric-invariant problem in boundary image matching. Supporting symmetric transformation is an important factor in boundary image matching to get more intuitive and more accurate matching results. However, the previous boundary image matching handled rotation transformation only without considering symmetric transformation. In this paper, we propose symmetric-invariant boundary image matching which supports the symmetric transformation as well as the rotation transformation. For this, we define the concept of image symmetry and formally prove that rotation-invariant matching of using a symmetric image always returns the same result for every symmetric angle. For efficient symmetric transformation, we also present how to efficiently extract the symmetric time-series from an image boundary. Finally, we formally prove that our symmetric-invariant matching produces the same result for two approaches: one is using the time-series extracted from the symmetric image; another is using the time-series directly obtained from the original image time-series by symmetric transformation. Experimental results show that the proposed symmetric-invariant boundary image matching obtains more accurate and intuitive results than the previous rotation-invariant boundary image matching. These results mean that our symmetric-invariant solution is an excellent approach that solves the image symmetry problem in time-series domain.

• The KIPS Transactions:PartB
• /
• v.11B no.4
• /
• pp.403-410
• /
• 2004

Recently, studies of image analysis, as the preprocessing stage for medical image analysis or image retrieval, are actively carried out. This paper intends to propose a way of utilizing color components for image retrieval. For image retrieval, it is based on color components, and for analysis of color, CLCM (Color Level Co-occurrence Matrix) and statistical techniques are used. CLCM proposed in this paper is to project color components on 3D space through geometric rotate transform and then, to interpret distribution that is made from the spatial relationship. CLCM is 2D histogram that is made in color model, which is created through geometric rotate transform of a color model. In order to analyze it, a statistical technique is used. Like CLCM, GLCM (Gray Level Co-occurrence Matrix)[1] and Invariant Moment [2,3] use 2D distribution chart, which use basic statistical techniques in order to interpret 2D data. However, even though GLCM and Invariant Moment are optimized in each domain, it is impossible to perfectly interpret irregular data available on the spatial coordinates. That is, GLCM and Invariant Moment use only the basic statistical techniques so reliability of the extracted features is low. In order to interpret the spatial relationship and weight of data, this study has used Principal Component Analysis [4,5] that is used in multivariate statistics. In order to increase accuracy of data, it has proposed a way to project color components on 3D space, to rotate it and then, to extract features of data from all angles.