• Journal of KIISE:Software and Applications
• /
• v.31 no.1
• /
• pp.1-7
• /
• 2004

For the handwritten digit recognition, this paper Proposes a hierarchical Gator features extraction method and a Bayesian network for them. Proposed Gator features are able to represent hierarchically different level information and Bayesian network is constructed to represent hierarchically structured dependencies among these Gator features. In order to extract such features, we define Gabor filters level by level and choose optimal Gabor filters by using Fisher's Linear Discriminant measure. Hierarchical Gator features are extracted by optimal Gabor filters and represent more localized information in the lower level. Proposed methods were successfully applied to handwritten digit recognition with well-known naive Bayesian classifier, k-nearest neighbor classifier. and backpropagation neural network and showed good performance.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2007.10a
• /
• pp.418-421
• /
• 2007

In this paper, we proposed a hierarchical nearest-neighbor searching method for deciding fitness of a clustered segment. It is difficult to distinguish the difference between correct spots and atypical noisy spots in footprint patterns. Therefore we could not completely remove unsuitable noisy spots from binarized image in image preprocessing stage or clustering stage. As a preprocessing stage for recognition of insect footprints, this method decides whether a segment is suitable or not, using degree of clustered segment fitness, and then unsuitable segments are eliminated from patterns. Removing unsuitable segments can improve performance of feature extraction for recognition of inset footprints.

• The Journal of the Korea Contents Association
• /
• v.10 no.1
• /
• pp.10-18
• /
• 2010

Due to a large number of classes in Hangul character recognition, it is usual to use the six-type preclassification stage. After the preclassification, the first consonent, vowel, and last consonent can be classified separately. Though each of three components has a few of classes, classification errors occurs often due to shape similarity such as 'ㅔ' and 'ㅖ'. So this paper proposes a hierarchical recognition method which adopts multi-stage tree structures for each of 6-types. In addition, to reduce the interference among three components, the method uses the recognition results of first consonents and vowel as features of vowel classifier. The recognition accuracy for the test set of PHD08 database was 98.96%.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.17 no.2
• /
• pp.15-23
• /
• 2017

The person Re-identification is the most challenging part of computer vision due to the significant changes in human pose and background clutter with occlusions. The picture from non-overlapping cameras enhance the difficulty to distinguish some person from the other. To reach a better performance match, most methods use feature selection and distance metrics separately to get discriminative representations and proper distance to describe the similarity between person and kind of ignoring some significant features. This situation has encouraged us to consider a novel method to deal with this problem. In this paper, we proposed an enhanced recurrent neural network with three-tier hierarchical network for person re-identification. Specifically, the proposed recurrent neural network (RNN) model contain an iterative expectation maximum (EM) algorithm and three-tier Hierarchical network to jointly learn both the discriminative features and metrics distance. The iterative EM algorithm can fully use of the feature extraction ability of convolutional neural network (CNN) which is in series before the RNN. By unsupervised learning, the EM framework can change the labels of the patches and train larger datasets. Through the three-tier hierarchical network, the convolutional neural network, recurrent network and pooling layer can jointly be a feature extractor to better train the network. The experimental result shows that comparing with other researchers' approaches in this field, this method also can get a competitive accuracy. The influence of different component of this method will be analyzed and evaluated in the future research.

• Proceedings of the Korean Information Science Society Conference
• /
• 2008.06c
• /
• pp.498-502
• /
• 2008

본 논문에서는 스케치 보정을 위한 계층적 지식 기반 잡음 제거 방법을 제안한다. 제안하는 잡음 제거 방법은 방향 정보, 후보 영역간의 내적, 갈고리 잡음영역 검출이라는 세 개의 계층적 휴리스틱(heuristic) 방법으로 구성된다. 첫 번째 단계에서 방향정보를 이용하여 특징점 후보들이 검출되고, 두 번째 단계에서는 각 후보들 사이의 벡터 간 내적을 이용하여 부적절한 후보들이 제거되며, 세 번째 단계에서는 갈고리모양의 잡음영역을 검출하여 근거리에 모여있는 특징점들을 병합한다. 실험을 통해 제안하는 방법이 잡음에 민감한 실제 응용 환경에 적합하며 효율적임을 보였다.

• The KIPS Transactions:PartB
• /
• v.17B no.2
• /
• pp.139-148
• /
• 2010

This paper shows a reinforcement post-processing method and feedback algorithm for improvement of assigning method in classification. Especially, we focused on complex documents that are generally considered to be hard to classify. A basis factors in traditional classification system are training methodology, classification models and features of documents. The classification problem of the documents containing shared features and multiple meanings, should be deeply mined or analyzed than general formatted data. To address the problems of these document, we proposed a method to expand classification scheme using decision boundary detected automatically in our previous studies. The assigning method that a document simply decides to the top ranked category, is a main factor that we focus on. In this paper, we propose a post-processing method and feedback algorithm to analyze the relevance of ranked list. In experiments, we applied our post-processing method and one time feedback algorithm to complex documents. The experimental results show that our system does not need to change the classification algorithm itself to improve the accuracy and flexibility.

• The KIPS Transactions:PartB
• /
• v.17B no.3
• /
• pp.207-214
• /
• 2010

A method of the region extraction and recognition of a traffic light and speed sign board in the real road environment is proposed. Traffic light was recognized by using brightness and color information based on HSI color model. Speed sign board was extracted by measuring red intensity from the HSI color information We improve the recognition rate by performing an incline compensation of the speed sign for directions clockwise and counterclockwise. The proposed algorithm shows a robust recognition rate in the image sequence which includes traffic light and speed sign board.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.95-96
• /
• 2021

최근 미디어의 발전으로 빠른 속도로 많은 양의 사람들의 얼굴이 포함된 사진, 동영상들이 인터넷에 업로드 되고 있다. 이러한 현상에 맞춰 인공지능을 활용한 얼굴 인식 기술의 놀라운 발전이 있었으나, 대규모 데이터셋에서 임의의 인물을 검색하는 경우에서는 연산량과 저장공간의 부담이 존재한다. 특히, 인터넷에 존재하는 수많은 불법 촬영물에서 피해자를 정확하고 신속하게 검색하기 위해서는 효율적인 얼굴 검색 시스템이 필요하다. 따라서, 본 논문은 얼굴 특징 추출과 클러스터링을 활용하여 방대한 양의 불법 촬영물 셋에서 피해자 동영상을 효율적으로 검색할 수 있는 기법을 제안한다. 불법 촬영물 동영상 검색 실험 환경을 만들기 위해 YouTube Faces [1] 데이터셋으로 유사 동영상 셋을 만들고 이 환경에서 실험을 진행한다. 얼굴 특징 추출 모델은 ResNet100 네트워크를 CosFace 손실함수와 Glint360K 데이터셋으로 학습시킨 모델 [2]을 사용한다. 추출된 얼굴 특징들을 HAC(Hierarchical Agglomerative Clustering) 알고리즘으로 클러스터링 한 후, 클러스터 대푯값을 통해 얼굴 검색 실험을 했을 때의 실험 결과를 분석한다.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.14 no.1
• /
• pp.131-137
• /
• 2014

Although illegal reproduction of cartoon books are prevalent with rapid growth of webhard services and smartphone use, fingerprinting technology for product identification, as seen in music and videos, has not been developed yet. This leads to indiscriminate illegal reproduction of cartoon books, causing great amount of copyright damages from copyright infringement of the rightful owners. The copyright R&D project granted from the Korea Copyright Permission (Project Title: Identification and Copy Protection Technology of Bookscaned Text/Comic Books) has been carried out in order to develop technology to effectively identify illegal reproduction and distribution of scanned cartoon books. The developed technology will contribute to increase of royalty payments and robust ecosystem of cartoon book markets. The study is to propose an enhanced implementation model for identification of scanned cartoon books on the basis of hierarchical symmetric difference feature algorithms adopted from existing feature extraction algorithms for video.

• The Journal of Korean Institute of Information Technology
• /
• v.17 no.8
• /
• pp.11-19
• /
• 2019

Against advanced image-related crimes, a high level of digital forensic methods is required. However, feature-based methods are difficult to respond to new device features by utilizing human-designed features, and deep learning-based methods should improve accuracy. This paper proposes a deep learning model to identify camera models based on DenseNet, the recent technology in the deep learning model field. To extract camera sensor features, a HPF feature extraction filter was applied. For camera model identification, we modified the number of hierarchical iterations and eliminated the Bottleneck layer and compression processing used to reduce computation. The proposed model was analyzed using the Dresden database and achieved an accuracy of 99.65% for 14 camera models. We achieved higher accuracy than previous studies and overcome their disadvantages with low accuracy for the same manufacturer.