• Phonetics and Speech Sciences
• /
• v.9 no.4
• /
• pp.27-34
• /
• 2017

The purpose of this work is to examine the English and Korean voiceless plosives from the Buckeye[15] and Seoul[16] corpus in terms of their static spectral characteristics. The plosives were automatically extracted by a Praat script. In order to estimate the percent correctness in the classification of the plosives, discriminant analyses were performed whose trainings were based on four spectral moments, i.e. the center of gravity, variance, skewness and kurtosis as suggested in [6]. Another set of discriminant analyses were performed based on the spectral tilts. In the last set of analyeses, the spectral moments and tilts were both used in the training. Results showed that the correct classification rate did not exceed around 65% in the best case, which suggested that phonetic cues other than the release burst would be necessary including the dynamic spectral aspects and vowel-onset cues.

• Journal of Information Processing Systems
• /
• v.16 no.5
• /
• pp.1105-1112
• /
• 2020

Gait recognition, as a promising biometric, can be used in video-based surveillance and other security systems. However, due to the complexity of leg movement and the difference of external sampling conditions, gait recognition still faces many problems to be addressed. In this paper, an improved convolutional neural network (CNN) based on Gabor filter is therefore proposed to achieve gait recognition. Firstly, a gait feature extraction layer based on Gabor filter is inserted into the traditional CNNs, which is used to extract gait features from gait silhouette images. Then, in the process of gait classification, using the output of CNN as input, we utilize metric learning techniques to calculate distance between two gaits and achieve gait classification by k-nearest neighbors classifiers. Finally, several experiments are conducted on two open-accessed gait datasets and demonstrate that our method reaches state-of-the-art performances in terms of correct recognition rate on the OULP and CASIA-B datasets.

• Proceedings of the IEEK Conference
• /
• summer
• /
• pp.371-375
• /
• 2004

Fault diagnosis is one of the most studied problems in process engineering. Recently, great research interest has been devoted to approaches that use classification methods to detect faults. This paper presents an application of a newly developed classification method - support vector machines - for fault diagnosis in an industrial case. A real set of operation data of a motor pump was used to train and test the support vector machines. The experiment results show that the support vector machines give higher correct detection rate of faults in comparison to rule-based diagnostics. In addition, the studied method can work with fewer training instances, what is important for online diagnostics.

• Journal of Korean Institute of Industrial Engineers
• /
• v.17 no.2
• /
• pp.27-37
• /
• 1991

A recognition algorithm for Hangul is developed by structural analysis to Hangul in this theses. Four major procedures are proposed : preprocessing, type classification, separation of consonant and vowel, recognition. In the preprocessing procedure, the thinning algorithm proposed by CHEN & HSU is applied. In the type classification procedure, thinned Hangul image is classified into one of six formal types. In the separation of consonant and vowel procedure, starting from branch-points which are existed in a vowel, character elements are separated by means of tracing branch-point pixel by pixel and comparison with proposed templates. In the same time, the vowels are recognized. In the recognition procedure, consonants are extracted from the separated Hangul character and recognized by modified Crossing method. Recognized characters are converted into KS-5601-1989 codes. The experiments show that correct recognition rate is about 80%-90% and recognition speed is about 2-3 character persecond in three types of different input data on computer with 80386 microprocessor.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.9
• /
• pp.3243-3257
• /
• 2021

Deep neural networks provide excellent performance in pattern recognition, audio classification, and image recognition. It is important that they accurately recognize input data, particularly when they are used in autonomous vehicles or for medical services. In this study, we propose a data correction method for increasing the accuracy of an unknown classifier by modifying the input data without changing the classifier. This method modifies the input data slightly so that the unknown classifier will correctly recognize the input data. It is an ensemble method that has the characteristic of transferability to an unknown classifier by generating corrected data that are correctly recognized by several classifiers that are known in advance. We tested our method using MNIST and CIFAR-10 as experimental data. The experimental results exhibit that the accuracy of the unknown classifier is a 100% correct recognition rate owing to the data correction generated by the proposed method, which minimizes data distortion to maintain the data's recognizability by humans.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2015.10a
• /
• pp.1045-1046
• /
• 2015

When using the Support Vector Machine in the online signature verification, SVM kernel function should be chosen to use non-linear SVM and the constant parameters in the kernel functions should be adjusted to appropriate values to reduce the error rate of signature verification. Non-linear SVM which is built on a strong mathematical basis shows better performance of classification with the higher discrimination power. However, choosing the kernel function and adjusting constant parameter values depend on the heuristics of the problem domain. In the signature verification, this paper deals with the problems of selecting the correct kernel function and constant parameters' values, and shows the kernel function and coefficient parameter's values with the minimum error rate. As a result of this research, we expect the average error rate to be less than 1%.

• Journal of the Institute of Convergence Signal Processing
• /
• v.2 no.4
• /
• pp.22-30
• /
• 2001

In this paper we propose three classification algorithms to classify breast tumors that occur in duct into Benign, DCIS(ductal carcinoma in situ) NOS(invasive ductal carcinoma) The general approach for a creating classifier is composed of 2 steps: feature extraction and classification Above all feature extraction for a good classifier is very significance, because the classification performance depends on the extracted features, Therefore in the feature extraction step, we extracted morphology features describing the size of nuclei and texture features The internal structures of the tumor are reflected from wavelet transformed images with 10$\times$ and 40$\times$ magnification. Pariticulary to find the correlation between correct classification rates and wavelet depths we applied 1, 2, 3 and 4-level wavelet transforms to the images and extracted texture feature from the transformed images The morphology features used are area, perimeter, width of X axis width of Y axis and circularity The texture features used are entropy energy contrast and homogeneity. In the classification step, we created three classifiers from each of extracted features using discriminant analysis The first classifier was made by morphology features. The second and the third classifiers were made by texture features of wavelet transformed images with 10$\times$ and 40$\times$ magnification. Finally we analyzed and compared the correct classification rate of the three classifiers. In this study, we found that the best classifier was made by texture features of 3-level wavelet transformed images.

• Journal of Applied Biological Chemistry
• /
• v.58 no.4
• /
• pp.345-348
• /
• 2015

Dried Rehmannia Radix Preparat, produced in different geographic origins of Korea and China, were investigated for the trace element contents using X-ray fluorescence spectrometry. When the estimated data for 35 kind of elements including K, P, S, Cl, Si, Al, Fe, Sn, etc on 143 samples were analyzed by canonical discrimination technique, the average classification correct rate was 92.3% and the significance was less than 0.0001. In case of reducing the number of elements for statistic analysis from 35 to 8 and to 3, the correct rates were also reduced to 88.1% and to 84.6%. The correct rate for Chinese samples ranged from 94.6-96.0%, while those for Korean ranged from 72.5-89.9%.

• Journal of Korea Society of Industrial Information Systems
• /
• v.27 no.2
• /
• pp.11-24
• /
• 2022

For the questions asked for LMS-based online evaluation the professor directly set exam questions, or use the automatic question-taking method according to the level of difficulty using the question bank divided by category. Among them, it is important to manage the difficulty of questions in an objective and efficient way, above all, in the automatic question-taking method according to difficulty. Because the questions presented to the evaluators may be different. In this paper, we propose an difficulty re-adjustment algorithm that considers not only the correct rate of a problem but also the time taken to solve the problem. For this, a logistic regression classification algorithm was used of machine learning, and a reference threshold was set based on the predicted probability value of the learning model and used to readjust the difficulty of each item. As a result, it was confirmed that there were many changes in the difficulty of each item that depended only on the existing correct rate. Also, as a result of performing group evaluation using the adjustment difficulty problem, it was confirmed that the average score improved in most groups compared to the difficulty problem based on the percentage of correct answers.

• Journal of Digital Convergence
• /
• v.11 no.9
• /
• pp.289-299
• /
• 2013

The aim of this study was to enhance the NHI claims data-based tuberculosis classification rule of KCDC(Korea centers for disease control & prevention) for an effective TB surveillance system. 8,118 cases, 10% samples of 81,199 TB cases from NHI claims data during 2009, were subject to the Medical Record Survey about whether they are real TB patients. The final study population was 7,132 cases whose medical records were surveyed. The decision tree model was evaluated as the most superior TB patients detection model. This model required the main independent variables of age, the number of anti-tuberculosis drugs, types of medical institution, tuberculosis tests, prescription days, types of TB. This model had sensitivity of 90.6%, PPV of 96.1%, and correct classification rate of 93.8%, which was better than KCDC's TB detection model with two or more NHI claims for TB and TB drugs(sensitivity of 82.6%, PPV of 95%, and correct classification rate of 80%).