http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
인쇄체 한글인식을 위한 MLP인식기의 인식결과 재추정 : Softmax 타당성 연구
임길택,김기석 경주대학교 정보전자기술연구소 2007 情報電子技術論叢 Vol.6 No.-
In this paper, we have studied on the feasibility of softmax method for MLP classifier which had been developed for the recognition of Type 1 machine printed Hangul character. When an MLP has been employed as a classifier, the softmax method has well known to be a proper method to obtain class a posteriori probability of the input character class. The recognition of postal address images is indispensable for the automatic sorting of postal envelopes. The process of the address image recognition is composed of three steps; address image preprocessing, character recognition, and finally address interpretation. The last address interpretation step is highly influenced by individual character recognition characteristics. For better envelope sorting rate, the character classifier should produce and forward proper results to the address interpretor. We have tested softmax and original method for reestimation of MLP classifier output values. To find out which method is more proper, we have utilized the character images of the real postal envelopes from the sorters in the post office. Through the experiments, we have seen that the original method produces better outputs for the address interpretor in terms of error and rejection for individual characters and non-characters.
효율적인 영상 이진화를 위한 적응적 Water Flow Model에 관한 연구
임길택,오현화 경주대학교 정보전자기술연구소 2004 情報電子技術論叢 Vol.3 No.-
In this paper, we propose an improved method to overcome the drawbacks of the existing water flow model based method for document image binarization. The proposed method defines the region of interest (ROI) around character areas to restrict rainfall onto the terrain surface. The amount of water required to fill a local valley is automatically determined according to the properties of the valley, thereby eliminating an iterative rainfall procedure which is very time consuming process in the existing method. Finally, to threshold the amount of water filled, a local adaptive method is applied to each pond. In the experiments on three document images, the proposed method showed better performances than the existing water flow model based method in terms of both quality and speed.
임길택 경주대학교 정보전자기술연구소 2006 情報電子技術論叢 Vol.5 No.-
In this paper, we have investigated the pattern classification performance of two feature vectors, PCA and RBF which have been frequently utilized for the classification of image patterns. The patterns considered were handwritten numeral images. We first have extracted direction angle vectors from image patterns. Then, PCA and RBF feature vectors have been obtained from those direction an인e vectors. The discrimination capability of two feature vectors has been analyzed in terms of the class separability measure Q which was defined in this paper. The class separability Q is obtained by calculating two kinds of distances, average pattern distance in the same class and average class distance between classes. Another measure for the feature vector's discrimination capability is based on the recognition rates of two classifiers, Bayes classifier and LMSE classifier. Experiments had been conducted on the handwritten numeral image database collected by Concordia university in Canada.
임길택 경주대학교 정보전자기술연구소 2008 情報電子技術論叢 Vol.7 No.-
In this paper, we have studied on the feasibility of the distance transformation of binary character string image. The existing nonlinear segmentation methods of individual character image from character string image utilize brightness values of a gray level image or different weight values which are assigned to transition path according to search direction. Unlike those methods, we adopt a systematic method using distance transformation to search a nonlinear segmentation line. To investigate the distance transformation method for nonlinear character segmentation, we include several experiments such as the comparison of resulting distance images by two transformation methods of Shih and Borgefors, the analysis of local distance peak of distance maps and distance value histogram. The experimental results have shown that the proposed method based on distance transformation is promising for nonlinear character segmentation of a binary character string image.
임길택 경주대학교 정보전자기술연구소 2005 情報電子技術論叢 Vol.4 No.-
This paper describes two combination methods of two statistical classifiers which are developed for the recognition of Handwritten Hangul characters. The goal of this research is to design a handwritten Korean character recognizer as a preprocessor of address interpretation for Korean mail pieces. For automatic mail processing, a typical process consists of several steps: mail image acquisition interpretation. To make address interpretation easier and more efficient, the character recognizer should retain the following three characteristics: reliable recognition scores indicating probability, high speed, and naturally acceptable cumulative recognition rates For the elementary classifiers, we have adopted two statistical classifiers, minimum distance classifier (MDC) and subspace classifier (SC), To satisfy the first characteristic and proposed two methods combining the two classifiers to meet the second and third ones. In the first combination method (CM1), the MDC classifier makes candidates for an input character image and forwards them to the SC which reorders the MDC's candidates. In the second combination method (CM2), we introduce the confusion matrix retaining information of the MDC's decision properties. In the CM2, the MDC also generates some candidates and forwards them to the SC which reorders the MDC's candidates as well as those from the confusion matrix. The superiority of the proposed combination methods has been proven through experiments done with the Handwritten Hangul character image database called PE92.
필기체 주소인식을 위한 MLP 기반 숫자인식기의 구현 방법
임길택 慶州大學校 2005 論文集 Vol.18 No.2
In the past several decades, a wide variety of research works have been conducted to develop the classification system of handwritten numerals. Most of them assumed that input characters are all valid numerals and did not give much attention to non-numeral inputs. If a numeral classifier constructed on only numeral data set is utilized for address recognition, we cannot expect what the result is when it tries to recognize non-numeral inputs. To build address recognition system for a Korean letter envelope, the designer should consider various kinds of inputs such as English characters, Chinese characters, Korean characters, and mis-segmented characters resulted from erroneous character segmentation process. In this paper, we investigate the properties of neural network based numeral classifiers to be used for Korean address recognition. Three types of errors are newly defined to inspect the characteristics of classifiers. Utilizing those error types, we present a promising neural network based method for the numeral classification.