• 정보처리학회논문지B
• /
• 제11B권6호
• /
• pp.667-674
• /
• 2004

이미지에서 문자 추출은 영상을 이해하기 위한 가장 기초적이고 중요한 문제이다. 본 논문에서는 문자의 획 특징을 이용하는 통계적인 방법으로 문자 영역을 검증하는 방법을 제안한다. 제안하는 방법은 $16\times16$ 크기의 텍스트와 비텍스트 이미지를 웨이블릿(wavelet) 변환하여 문자의 획과 방향성을 표현하는 35차원의 특징을 추출한다. 추출된 특징 중 변별력이 높은 특징만을 선택하여 SVM(Support Vector Machine) 분류기를 구성한다. 분류기론 이용하여 $16\times16$크기의 윈도우로 검증 영역을 스캔하면서, 각각의 윈도우를 텍스트와 비텍스트로 분류하고 최종적으로 검증 영역의 텍스트 여부를 결정한다. 제안한 방법을 적용함으로써 텍스트와 유사하여 구별하기 어려운 비텍스트 영역을 검증할 수 있었다.

• 한국멀티미디어학회논문지
• /
• 제10권6호
• /
• pp.699-706
• /
• 2007

This article is an approach to improving the clustering of Web images by using high-level semantic features from text information relevant to Web images as well as low-level visual features of image itself. These high-level text features can be obtained from image URLs and file names, page titles, hyperlinks, and surrounding text. As a clustering algorithm, a self-organizing map (SOM) proposed by Kohonen is used. To evaluate the clustering efficiencies of SOMs, we propose a simple but effective measure indicating the accumulativeness of same class images and the perplexities of class distributions. Our approach is to advance the existing measures through defining and using new measures accumulativeness on the most superior clustering node and concentricity to evaluate clustering efficiencies of SOMs. The experimental results show that the high-level text features are more useful in SOM-based Web image clustering.

• 정보처리학회논문지B
• /
• 제13B권2호
• /
• pp.121-126
• /
• 2006

본 논문에서는 웹 이미지의 분류 효과를 높이기 위해 이미지 자체에서 추출된 저수준의 비주얼 특징뿐만 아니라 이미지와 관련된 텍스트 정보로부터 나온 고수준 시맨틱 특징들을 이용하는 분류 방법을 제안한다. 이 고수준의 텍스트 특징들은 이미지 URL, 파일명, 페이지 타이틀, 하이퍼링크 및 이미지 주변 텍스트로부터 얻어진다. 분류 엔진으로는 Kohonen의 SOM(Self Organizing Map)을 사용한다. 고수준의 텍스트 특징들과 저수준의 비주얼 특징들을 동시에 사용하는 SOM 기반의 이미지 분류에서는 10개의 카테고리로부터 수집된 200개의 테스트 이미지들이 사용되었다. 분류 성능을 평가하기 위해 간단하면서도 새로운 두 가지 척도, 즉 동일 카테고리 이미지들의 산포 정도와 집적 정도를 나타내는 각각의 척도를 정의하고 사용하였다. 실험결과, SOM기반의 웹 이미지 분류에서는 고수준의 텍스트 특징들이 보다 유용한 것임이 밝혀졌다.

• 한국방송∙미디어공학회:학술대회논문집
• /
• 한국방송공학회 2009년도 IWAIT
• /
• pp.32-37
• /
• 2009

Recently, electronic mail(E-mail) is the most popular communication manner in our society. In such conventional environments, spam increasingly congested in Internet. In this paper, Chinese spam could be effectively detected using text and image features. Using text features, keywords and reference templates in Chinese mails are automatically selected using genetic algorithm(GA). In addition, spam containing a promotion image is also filtered out by detecting the text characters in images. Some experimental results are given to show the effectiveness of our proposed method.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제6권3호
• /
• pp.941-953
• /
• 2012

A novel and universal method of video image text detection is proposed. A coarse-to-fine text detection method is implemented. Firstly, the spectral clustering (SC) method is adopted to coarsely detect text regions based on the stationary wavelet transform (SWT). In order to make full use of the information, multi-parameters kernel function which combining the features similarity information and spatial adjacency information is employed in the SC method. Secondly, 28 dimension classifying features are proposed and support vector machine (SVM) is implemented to classify text regions with non-text regions. Experimental results on video images show the encouraging performance of the proposed algorithm and classifying features.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제6권4호
• /
• pp.1140-1152
• /
• 2012

A novel and universal method of video image text detection is proposed. A coarse-to-fine text detection method is implemented. Firstly, the spectral clustering (SC) method is adopted to coarsely detect text regions based on the stationary wavelet transform (SWT). In order to make full use of the information, multi-parameters kernel function which combining the features similarity information and spatial adjacency information is employed in the SC method. Secondly, 28 dimension classifying features are proposed and support vector machine (SVM) is implemented to classify text regions with non-text regions. Experimental results on video images show the encouraging performance of the proposed algorithm and classifying features.

• 전기학회논문지
• /
• 제58권5호
• /
• pp.1025-1034
• /
• 2009

In steel making production line, steel slabs are given a unique identification number. This identification number, Slab management number(SMN), gives information about the use of the slab. Identification of SMN has been done by humans for several years, but this is expensive and not accurate and it has been a heavy burden on the workers. Consequently, to improve efficiency, automatic recognition system is desirable. Generally, a recognition system consists of text localization, text extraction, character segmentation, and character recognition. For exact SMN identification, all the stage of the recognition system must be successful. In particular, the text localization is great important stage and difficult to process. However, because of many text-like patterns in a complex background and high fuzziness between the slab and background, directly extracting text region is difficult to process. If the slab region including SMN can be detected precisely, text localization algorithm will be able to be developed on the more simple method and the processing time of the overall recognition system will be reduced. This paper describes about the slab region localization using SIFT(Scale Invariant Feature Transform) features in the image. First, SIFT algorithm is applied the captured background and slab image, then features of two images are matched by Nearest Neighbor(NN) algorithm. However, correct matching rate can be low when two images are matched. Thus, to remove incorrect match between the features of two images, geometric locations of the matched two feature points are used. Finally, search rectangle method is performed in correct matching features, and then the top boundary and side boundaries of the slab region are determined. For this processes, we can reduce search region for extraction of SMN from the slab image. Most cases, to extract text region, search region is heuristically fixed [1][2]. However, the proposed algorithm is more analytic than other algorithms, because the search region is not fixed and the slab region is searched in the whole image. Experimental results show that the proposed algorithm has a good performance.

• 한국언어정보학회:학술대회논문집
• /
• 한국언어정보학회 2007년도 정기학술대회
• /
• pp.515-521
• /
• 2007

The text categorization is an important field for the automatic text information processing. Moreover, the authorship identification of a text can be treated as a special text categorization. This paper adopts the conceptual primitives' expression based on the Hierarchical Network of Concepts (HNC) theory, which can describe the words meaning in hierarchical symbols, in order to avoid the sparse data shortcoming that is aroused by the natural language surface features in text categorization. The KNN algorithm is used as computing classification element. Then, the experiment has been done on the Chinese text authorship identification. The experiment result gives out that the processing mode that is put forward in this paper achieves high correct rate, so it is feasible for the text authorship identification.

• Journal of Information Processing Systems
• /
• 제1권1호
• /
• pp.22-26
• /
• 2005

In this paper, we classified the objectionable texts into four rates according to their harmfulness and proposed the hierarchical text rating system for objectionable documents. Since the documents in the same category have similarities in used words, expressions and structure of the document, the text rating system, which uses a single classification model, has low accuracy. To solve this problem, we separate objectionable documents into several subsets by using their properties, and then classify the subsets hierarchically. The proposed system consists of three layers. In each layer, we select features using the chi-square statistics, and then the weight of the features, which is calculated by using the TF-IDF weighting scheme, is used as an input of the non-linear SVM classifier. By means of a hierarchical scheme using the different features and the different number of features in each layer, we can characterize the objectionability of documents more effectively and expect to improve the performance of the rating system. We compared the performance of the proposed system and performance of several text rating systems and experimental results show that the proposed system can archive an excellent classification performance.

• 스마트미디어저널
• /
• 제10권1호
• /
• pp.63-69
• /
• 2021

사람이 어떤 문장을 보고 그 문장에 대해 이해하는 것은 문장 안에서 주요한 단어를 이미지로 연상시켜 그 문장에 대해 이해한다. 이러한 연상과정을 컴퓨터가 할 수 있도록 하는 것을 text-to-image라고 한다. 기존 딥 러닝 기반 text-to-image 모델은 Convolutional Neural Network(CNN)-Long Short Term Memory(LSTM), bi-directional LSTM을 사용하여 텍스트의 특징을 추출하고, GAN에 입력으로 하여 이미지를 생성한다. 기존 text-to-image 모델은 텍스트 특징 추출에서 기본적인 임베딩을 사용하였으며, 여러 모듈을 사용하여 이미지를 생성하므로 학습 시간이 오래 걸린다. 따라서 본 연구에서는 자연어 처리분야에서 성능 향상을 보인 어텐션 메커니즘(Attention Mechanism)을 문장 임베딩에 사용하여 특징을 추출하고, 추출된 특징을 GAN에 입력하여 이미지를 생성하는 방법을 제안한다. 실험 결과 기존 연구에서 사용되는 모델보다 inception score가 높았으며 육안으로 판단하였을 때 입력된 문장에서 특징을 잘 표현하는 이미지를 생성하였다. 또한, 긴 문장이 입력되었을 때에도 문장을 잘 표현하는 이미지를 생성하였다.