• KIISE Transactions on Computing Practices
• /
• v.24 no.3
• /
• pp.151-156
• /
• 2018

In this paper, a method to classify objects in outdoor CCTV images using Convolutional Neural Network(CNN) and background subtraction is proposed. Object candidates are extracted using background subtraction and they are classified with CNN to detect objects in the image. At the end, computation complexity is highly reduced in comparison to other object detection algorithms. A database is constructed by filming alleys and playgrounds, places where crime occurs mainly. In experiments, different image sizes and experimental settings are tested to construct a best classifier detecting person. And the final classification accuracy became 80% for same camera data and 67.5% for a different camera.

• Korean Journal of Cognitive Science
• /
• v.2 no.1
• /
• pp.33-50
• /
• 1990

Recently, neural networks have been proposed as computaional models for hard prlblems that the brain appears to solve easily. This paper proposes a hierarchical network which practically recognizes printed Hangul characters based on the various psychological stueies. This system is composed of a type classification netwotk and six recognition networks. The former clessifier input character images into one of the six thper by their overall sturcture, and the latter further classify them into character code. Extperiments with most frequently used 990 printed hangul characters conform the superiority of the propsed system. After all, neural nework approach turns out to be very reasonable through a comparison with statistical classifier and an analysis of mis-classification and generalization capability.

• International Journal of Advanced Culture Technology
• /
• v.8 no.3
• /
• pp.316-324
• /
• 2020

It has experienced how difficult People with foreign language learning, it is to pronounce a new language different from the native language. The goal of various foreigners who want to learn Korean is to speak Korean as well as their native language to communicate smoothly. However, each native language's vocal habits also appear in Korean pronunciation, which prevents accurate information transmission. In this paper, the pronunciation of Chinese learners was compared with that of Korean. For comparison, the fundamental frequency and its variation of the speech signal were examined and the spectrogram was analyzed. The Formant frequencies known as the resonant frequency of the vocal tract were calculated. Based on these characteristics parameters, the classifier of the Supporting Vector Machine was found to classify the pronunciation of Koreans and the pronunciation of Chinese learners. In particular, the linguistic proposition was scientifically proved by examining the Korean pronunciation of /ㄹ/ that the Chinese people were not good at pronouncing.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.11 no.12
• /
• pp.6038-6053
• /
• 2017

In this study, we propose an automatic melody extraction algorithm using deep learning. In this algorithm, feature images, generated using the energy of frequency band, are extracted from polyphonic audio files and a deep learning technique, a convolutional neural network (CNN), is applied on the feature images. In the training data, a short frame of polyphonic music is labeled as a musical note and a classifier based on CNN is learned in order to determine a pitch value of a short frame of audio signal. We want to build a novel structure of melody extraction, thus the proposed algorithm has a simple structure and instead of using various signal processing techniques for melody extraction, we use only a CNN to find a melody from a polyphonic audio. Despite of simple structure, the promising results are obtained in the experiments. Compared with state-of-the-art algorithms, the proposed algorithm did not give the best result, but comparable results were obtained and we believe they could be improved with the appropriate training data. In this paper, melody extraction and the proposed algorithm are introduced first, and the proposed algorithm is then further explained in detail. Finally, we present our experiment and the comparison of results follows.

• Nuclear Engineering and Technology
• /
• v.47 no.2
• /
• pp.176-186
• /
• 2015

Pattern classifications have become important tools for fault diagnosis in nuclear power plants (NPP). However, it is often difficult to obtain training data under fault conditions to train a supervised classification model. By contrast, normal plant operating data can be easily made available through increased deployment of supervisory, control, and data acquisition systems. Such data can also be used to train classification models to improve the performance of fault diagnosis scheme. In this paper, a fault diagnosis scheme based on semisupervised classification (SSC) scheme is developed. In this scheme, new measurements collected from the plant are integrated with data observed under fault conditions to train the SSC models. The trained models are subsequently applied to new measurements for fault diagnosis. In comparison with supervised classifiers, the proposed scheme requires significantly fewer data collected under fault conditions to train the classifier. The developed scheme has been validated using different fault scenarios on a desktop NPP simulator as well as on a physical NPP simulator using a graph-based SSC algorithm. All the considered faults have been successfully diagnosed. The results have demonstrated that SSC is a promising tool for fault diagnosis in NPPs.

• Journal of Sensor Science and Technology
• /
• v.29 no.1
• /
• pp.19-26
• /
• 2020

Electrocardiogram (ECG) classification has become an essential task of modern day wearable devices, and can be used to detect cardiovascular diseases. State-of-the-art Artificial Intelligence (AI)-based ECG classifiers have been designed using various artificial neural networks (ANNs). Despite their high accuracy, ANNs require significant computational resources and power. Herein, three different ANNs have been compared: multilayer perceptron (MLP), convolutional neural network (CNN), and spiking neural network (SNN) only for the ECG classification. The ANN model has been developed in Python and Theano, trained on a central processing unit (CPU) platform, and deployed on a PYNQ-Z2 FPGA board to validate the model using a Jupyter notebook. Meanwhile, the hardware accelerator is designed with Overlay, which is a hardware library on PYNQ. For classification, the MIT-BIH dataset obtained from the Physionet library is used. The resulting ANN system can accurately classify four ECG types: normal, atrial premature contraction, left bundle branch block, and premature ventricular contraction. The performance of the ECG classifier models is evaluated based on accuracy and power. Among the three AI algorithms, the SNN requires the lowest power consumption of 0.226 W on-chip, followed by MLP (1.677 W), and CNN (2.266 W). However, the highest accuracy is achieved by the CNN (95%), followed by MLP (76%) and SNN (90%).

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.10 no.5
• /
• pp.956-961
• /
• 2006

Keystroke time intervals can be a discriminating feature in the verification and identification of computer users. This paper presents a comparison result obtained using several classification methods including k-NN (k-Nearest Neighbor), back-propagation neural networks, and Bayesian classification for keystroke identification. Performance of k-NN classification was best with small data samples available per user, while Bayesian classification was the most superior to others with large data samples per user. Thus, for web-based on-line identification of users, it seems to be appropriate to selectively use either k-NN or Bayesian method according to the number of keystroke samples accumulated by each user.

• The Korean Journal of Applied Statistics
• /
• v.26 no.1
• /
• pp.131-139
• /
• 2013

Support vector machines(SVM) may perform poorly in the presence of noise variables; in addition, it is difficult to identify the importance of each variable in the resulting classifier. A feature selection can improve the interpretability and the accuracy of SVM. Most existing studies concern feature selection in the linear SVM through penalty functions yielding sparse solutions. Note that one usually adopts nonlinear kernels for the accuracy of classification in practice. Hence feature selection is still desirable for nonlinear SVMs. In this paper, we compare the performances of nonlinear feature selection methods such as component selection and smoothing operator(COSSO) and kernel iterative feature extraction(KNIFE) on simulated and real data sets.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.2
• /
• pp.647-669
• /
• 2016

Multiple attribute for decision making including user preference will increase the complexity of route selection process. Various approaches have been proposed to solve the optimal route selection problem. In this paper, multi attribute decision making (MADM) algorithms such as Simple Additive Weighting (SAW), Weighted Product Method (WPM), Analytic Hierarchy Process (AHP) method and Total Order Preference by Similarity to the Ideal Solution (TOPSIS) methods have been proposed for acoustic signature based optimal route selection to facilitate user with better quality of service. The traffic density state conditions (very low, low, below medium, medium, above medium, high and very high) on the road segment is the occurrence and mixture weightings of traffic noise signals (Tyre, Engine, Air Turbulence, Exhaust, and Honks etc) is considered as one of the attribute in decision making process. The short-term spectral envelope features of the cumulative acoustic signals are extracted using Mel-Frequency Cepstral Coefficients (MFCC) and Adaptive Neuro-Fuzzy Classifier (ANFC) is used to model seven traffic density states. Simple point method and AHP has been used for calculation of weights of decision parameters. Numerical results show that WPM, AHP and TOPSIS provide similar performance.

• International Journal of Contents
• /
• v.11 no.2
• /
• pp.15-19
• /
• 2015

In this paper, we propose an efficient document layout analysis algorithm that includes table detection. Typical methods of document layout analysis use the height and gap between words or columns. To correspond to the various styles and sizes of documents, we propose an algorithm that uses the mean value of the distance transform representing thickness and compare with components in the local area. With this algorithm, we combine a table detection algorithm using the same feature as that of the text classifier. Table candidates, separators, and big components are isolated from the image using Connected Component Analysis (CCA) and distance transform. The key idea of text classification is that the characteristics of the text parallel components that have a similar thickness and height. In order to estimate local similarity, we detect a text region using an adaptive searching window size. An improved adaptive run-length smoothing algorithm (ARLSA) was proposed to create the proper boundary of a text zone and non-text zone. Results from experiments on the ICDAR2009 page segmentation competition test set and our dataset demonstrate the superiority of our dataset through f-measure comparison with other algorithms.