低域通過濾波器의 振幅應答特性의 改善

高昌林 慶北工業專門大學 1979 論文集 Vol.16 No.-

The magnitude approximation of low pass Butterworth is excellent at and neat ω=0, but poor at ω=??. And it possess excellent phase linearity. In this paper, by cascading second order low pass and band stop filter, the improvement of attenuation in the stop band is obtained. Experimental results are presented and discussed.

필기체 접촉 숫자들의 분할 기반 인식

김진호,고창림 경일대학교 산업기술연구소 1999 産業技術硏究所 論文集 Vol.6 No.-

본 논문에서는 필기체 접촉 숫자들의 효과적인 분할기반 인식 방안을 제안하였다. 먼저 접촉 숫자들을 숫자들간 연결획 해석 (ligature analysis) 을 통하여 여섯 가지 유형으로 분류하였다. 그리고 네 종류의 절단 후보점들을 찾은 다음 해당 후보 세그먼트들의 유형을 해석하고 최종 세그먼트 위치를 결정하였다. 제안된 방법은 인식 대상 세그먼트 조합의 개수를 최소화 시킬 뿐만 아니라 기존의 분할기법 인식에서 흔히 발생되는 분할에러를 최소화 시킬 수 있는 장점을 가지고 있다. 제안된 방법에 따라 인식 시스템을 구현하고 NIST 접촉 숫자 데이터베이스의 5000개 숫자를 대상으로 인식 실험해 본 결과 기존의 연구 결과들에 비해 보다 개선된 89.0%의 인식률을 얻을 수 있었다. This paper proposes an efficient segmentation based recognitive method for unconstrained handwritten touching digits. We classified touching digits into six types which are based on the ligature analysis. The candidate break points are sought and the final segmentation point of touching digit is decided through segment verification of candidate segment combinations. The main advantages of this method are that only reliable reduced number of candidates are used in the multiple hypothesis recognition, and segmentation error of traditional segmentation based approach is also decreased. To evaluate proposed method, we have experimented for the NIST database of two touching numerals with 5,000 digits and obtained the more improved results with 89.0% recognition rates than previous several reports.

ART 신경망을 이용한 패턴들의 집단화 과정

高昌林 慶一大學校 1998 論文集 Vol.15 No.1

In this paper, Two aspects of pattern clustering process using ART neural network are studied and discussed. First, the relations between vigilance parameter values and the number of clustering centers are examined. It can be concluded that the number of cluster centers are not linearly increased for the linear increase of vigilance parameter(from 0.1 to 0.9) in clustering procedure. Second, for given vigilance parameter(0.95) with the ordered sequence of input patterns, the related informations of pattern clustering process are examined. In this procedure, if any of the stored rpototype patterns is similar to input pattern, then the input pattern is included to that cluster center with the prototype pattern is modified to the input pattern for the matching. Otherwise new cluster center is formed and the input pattern is included to that cluster. Moreover the result of pattern clustering process are varied on the order of input pattern sequence. Experimental results are obtained and discussed.

영상분할을 위한 이완법의 초기확률

高昌林 慶一大學校 1995 論文集 Vol.12 No.2

Two approaches to the initial probability for the image segmentation using the relaxation process are studied. The first initial probabilities are defined based on the image histogram. In this case, the convergence rate of label probability is fast and an image is segmented to the roughly segmented regions. Therefore this method is suitable for the large size general image. Another initial probability is derived from the local feature of an image. Although the convergence rate of pixel labeling is relatively slow, an image is segmented to feature regions with low segmentation error. Experimental results show that it is very useful for the medical image segamentation with the low contrast and small size area of interest.

병렬 GHT 모델을 이용한 다중형태인식

金鎭浩,趙三模,高昌林 慶一大學校 1993 論文集 Vol.9 No.1

A parallel GHT model is introduced which can detect multiple shapes simultaneously with an efficient use of Hough space and modification of its accumulation method. In this method, we separated the Hough space into two and accumulated the evidences of reference points in two directions of up and down. The proposed model is implemented on the Transputer system and the resulted system proves to have simultaneous rotation-and scaling-invariant detection capability for four different shapes.

패턴벡터들간의 직교성에 따른 홉필드 신경망의 복원특성 : -잡음이 인가된 부분패턴의 경우- -For Noisy and Partial Test Pattern-

高昌林 慶一大學校 1999 論文集 Vol.16 No.5

Recall properties of the Hopfield neural network based on orthogonality among the exemplar pattern vectors with the partial and noisy test patterns are studied. The bitmap characters "KOREA" and four 10*10 bitmap patterns which are composed of 10 bits bar pattern combination are used as training patterns for the nonorthonormal and orthonormal case respectively. From the simulation results it is proved that nonorthonormal pattern case the recall of test pattern is not good regardless of shape and amount of noise to be added, and orthonormal pattern case exact recall of test pattern can be obtained regardless of shape and amount of nosie to be added.

弛緩法을 이용한 螢光眠底畵像의 Thresholding

高昌林 慶一大學校 1987 論文集 Vol.3 No.1

In this paper thresholding method via relaxation process is studied. In relaxation process initial probability for pixel classification is defined and the pixels are classified probabilistically. And then adjust the probabilities for each pixel based on its neighbors probabilities with light, dark, medium classes. When this adjustment process is iterated, the label probabilities are refined and the pixels are clustered in one of 3 classes so that thresholding becomes trivial with regardless of histogram distribution. Experimental results are presented and discussed on the ocular fundus fluorescein angiogram with low contrast and unimodal histogram distribution.

디지틀 화상처리에 의한 형광안저화상의 해석

高昌林 慶一大學校 1992 論文集 Vol.8 No.1

Digital image analysis for the ocular fluorescein angiogram with the low contrast and unimodal hisgram distribution and studied. The procedure are image segmentation, Correction, and the discrimination of the feature regions. Based on the relaxation process, thresholding by pixel clustering and the pixel classification through refined label probabilities are discussed. Moreover for the extraction of useful diagnostic information, correction via the registration of nonlinearly distorted images and discrimination of feature regions by its shape are studied.

差分演算子를 利用한 頂點과 溪谷點 檢出 : 畵像의 階層的 分析技法

高昌林 慶北工業專門大學 1984 論文集 Vol.21 No.-

A new technique on peak and valley detection by difference operator using restamping is developed. The gray level difference matrices are constructed from the two dimensional data, and are used for detection of local peaks and valleys. This hierarchical image search technique based on restamping is computationally efficient method. There are no parameters to adjustment of the process. Effects of noise are minimized without altering the peaks and valleys points. The results of the process can be used as a description of region properties.

오차 역전파 신경망의 수렴 특성과 문자 인식

高昌林 慶一大學校 1997 論文集 Vol.14 No.2

The convergence properties of back propagation algorithm for learning constant and momentum rate, and its application to character recognition are studied. Combination of 0.1, 0.5, 0.9 of learning rate and momentum constant, convergence characteristics are obtained and analyzed. And using selected values of these parameters with adding small random noise to input patterns, character recognition using back propagation neural network is tested and verified. At present the choice of these parameters are non-trival problem, the results in this paper will provide the useful guidelines for the choice of these parameters in the applications of back propagation neural networks.