• Journal of KIISE:Software and Applications
• /
• v.30 no.3_4
• /
• pp.339-346
• /
• 2003

This paper proposes a multiple classifier system having massive micro classifiers. The micro classifiers are trained by using a local set of training patterns. The k nearest neighboring training patterns of one training pattern comprise the local region for training a micro classifier. Each training pattern is incorporated with one or more micro classifiers. Two types of micro classifiers are adapted in this paper. SVM with linear kernel and SVM with RBF kernel. Classification is done by selecting the best micro classifier among the micro classifiers in vicinity of incoming test pattern. To measure the goodness of each micro classifier, the weighted sum of correctly classified training patterns in vicinity of the test pattern is used. Experiments have been done on Elena database. Results show that the proposed method gives better classification accuracy than any conventional classifiers like SVM, k-NN and the conventional classifier combination/selection scheme.

• Journal of Information Processing Systems
• /
• v.13 no.4
• /
• pp.928-940
• /
• 2017

State-of-the-art speaker recognition systems may work better for the English language. However, if the same system is used for recognizing those who speak different languages, the systems may yield a poor performance. In this work, the decisions of a Gaussian mixture model-universal background model (GMM-UBM) and a learning vector quantization (LVQ) are combined to improve the recognition performance of a multilingual speaker identification system. The difference between these classifiers is in their modeling techniques. The former one is based on probabilistic approach and the latter one is based on the fine-tuning of neurons. Since the approaches are different, each modeling technique identifies different sets of speakers for the same database set. Therefore, the decisions of the classifiers may be used to improve the performance. In this study, multitaper mel-frequency cepstral coefficients (MFCCs) are used as the features and the monolingual and cross-lingual speaker identification studies are conducted using NIST-2003 and our own database. The experimental results show that the combined system improves the performance by nearly 10% compared with that of the individual classifier.

• Korean Journal of Fisheries and Aquatic Sciences
• /
• v.50 no.2
• /
• pp.195-206
• /
• 2017

Joint time-frequency images of the broadband echo signals of six fish species were obtained using the smoothed pseudo-Wigner-Ville distribution in controlled environments. Affine transformation and principal component analysis were used to obtain eigenimages that provided species-specific acoustic features for each of the six fish species. The echo images of an unknown fish species, acquired in real time and in a fully automated fashion, were identified by finding the smallest Euclidean or Mahalanobis distance between each combination of weight matrices of the test image of the fish species to be identified and of the eigenimage classes of each of six fish species in the training set. The experimental results showed that the Mahalanobis classifier performed better than the Euclidean classifier in identifying both single- and mixed-species groups of all species assessed.

• Journal of Institute of Control, Robotics and Systems
• /
• v.8 no.10
• /
• pp.827-832
• /
• 2002

We propose a prediction method of the pollutant and a synchronous classification of the current state of SOx emission in the power plant. We use the auto-regressive with exogeneous (ARX) model as a predictor of SOx emission and use a radial basis function network (RBFN) as a pattem classifier. The ARX modeling scheme is implemented using recursive least squares (RLS) method to update the model parameters adaptively. The capability of SOx emission monitoring is utilized with the application of the RBFN classifier. Experimental results show that the ARX model can predict the SOx emission concentration well and ARX modeling parameters can be a good feature for the state monitoring. in addition, its validity has been verified through the power spectrum analysis. Consequently, the RBFN classifier in combination with ARX model is shown to be quite adequate for monitoring the state of SOx emission.

• Journal of Biomedical Engineering Research
• /
• v.17 no.3
• /
• pp.379-386
• /
• 1996

This paper describes an approach for classifying myoelectric patterns using a multilayer perceptrons (MLP's) and hidden Markov models (HMM's) hybrid classifier. The dynamic aspects of EMG are important for tasks such as continuous prosthetic control or vari- ous time length EMG signal recognition, which have not been successfully mastered by the most neural approaches. It is known that the hidden Markov model (HMM) is suitable for modeling temporal patterns. In contrasts the multilayer feedforward networks are suitable for static patterns. Ank a lot of investigators have shown that the HMM's to be an excellent tool for handling the dynamical problems. Considering these facts, we suggest the combination of MLP and HMM algorithms that might lead to further improved EMG recognition systems.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.37 no.2A
• /
• pp.95-104
• /
• 2012

Pattern recognition using ensemble classifiers is composed of strong classifier which consists of many weak classifiers. In this paper, we used feature extraction to organize strong classifier using static camera sequence. The strong classifier is made of weak classifiers which considers environmental factors. So the strong classifier overcomes environmental effect. Proposed method uses binary foreground image by frame difference method and the boosting is used to train crowdedness model and recognize crowdedness using features. Combination of weak classifiers makes strong ensemble classifier. The classifier could make use of potential features from the environment such as shadow and reflection. We tested the proposed system with road sequence and subway platform sequence which are included in "AVSS 2007" sequence. The result shows good accuracy and efficiency on complex environment.

• Journal of IKEEE
• /
• v.18 no.4
• /
• pp.536-543
• /
• 2014

Terrain classification is still a challenging issue in image processing, especially with high resolution satellite images. The well-known obstacles include low accuracy in the detection of targets, especially for the case of man-made structures, such as buildings and roads. In this paper, we present an efficient approach to classify and detect building footprints, foliage, grass and road from high resolution grayscale satellite images. Our contribution is to build a strong classifier using AdaBoost based on a combination of co-occurrence and Haar-like features. We expect that the inclusion of Harr-like feature improves the classification performance of the man-made structures, since Haar-like feature is extracted from corner features and rectangle features. Also, the AdaBoost algorithm selects only critical features and generates an extremely efficient classifier. Experimental result indicates that the classification accuracy of AdaBoost classifier is much higher than that of the conventional classifier using back propagation algorithm. Also, the inclusion of Harr-like feature significantly improves the classification accuracy. The accuracy of the proposed method is 98.4% for the target detection and 92.8% for the classification on high resolution satellite images.

• Journal of KIISE:Software and Applications
• /
• v.31 no.12
• /
• pp.1684-1690
• /
• 2004

This paper presents a method for classifier selection that uses distribution information of the training samples in a small region surrounding a sample. The conventional DCS-LA(Dynamic Classifier Selection - Local Accuracy) selects a classifier dynamically by comparing the local accuracy of each classifier at the test time, which inevitably requires long classification time. On the other hand, in the proposed approach, the best classifier in a local region is stored in the FSA(Feature Space Attribute) table during the training time, and the test is done by just referring to the table. Therefore, this approach enables fast classification because classification is not needed during test. Two feature space attributes are used entropy and density of k training samples around each sample. Each sample in the feature space is mapped into a point in the attribute space made by two attributes. The attribute space is divided into regular rectangular cells in which the local accuracy of each classifier is appended. The cells with associated local accuracy comprise the FSA table. During test, when a test sample is applied, the cell to which the test sample belongs is determined first by calculating the two attributes, and then, the most accurate classifier is chosen from the FSA table. To show the effectiveness of the proposed algorithm, it is compared with the conventional DCS -LA using the Elena database. The experiments show that the accuracy of the proposed algorithm is almost same as DCS-LA, but the classification time is about four times faster than that.

• International Journal of Control, Automation, and Systems
• /
• v.3 no.4
• /
• pp.571-579
• /
• 2005

In this paper, we proposed an algorithm for arrhythmia classification, which is associated with the reduction of feature dimensions by linear discriminant analysis (LDA) and a support vector machine (SVM) based classifier. Seventeen original input features were extracted from preprocessed signals by wavelet transform, and attempts were then made to reduce these to 4 features, the linear combination of original features, by LDA. The performance of the SVM classifier with reduced features by LDA showed higher than with that by principal component analysis (PCA) and even with original features. For a cross-validation procedure, this SVM classifier was compared with Multilayer Perceptrons (MLP) and Fuzzy Inference System (FIS) classifiers. When all classifiers used the same reduced features, the overall performance of the SVM classifier was comprehensively superior to all others. Especially, the accuracy of discrimination of normal sinus rhythm (NSR), arterial premature contraction (APC), supraventricular tachycardia (SVT), premature ventricular contraction (PVC), ventricular tachycardia (VT) and ventricular fibrillation (VF) were $99.307\%,\;99.274\%,\;99.854\%,\;98.344\%,\;99.441\%\;and\;99.883\%$, respectively. And, even with smaller learning data, the SVM classifier offered better performance than the MLP classifier.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.7 no.11
• /
• pp.2720-2736
• /
• 2013

Multi-view face detection has become an active area for research in the last few years. In this paper, a novel multi-view human face detection algorithm based on improved real Adaboost is presented. Real Adaboost algorithm is improved by weighted combination of weak classifiers and the approximately best combination coefficients are obtained. After that, we proved that the function of sample weight adjusting method and weak classifier training method is to guarantee the independence of weak classifiers. A coarse-to-fine hierarchical face detector combining the high efficiency of Haar feature with pose estimation phase based on our real Adaboost algorithm is proposed. This algorithm reduces training time cost greatly compared with classical real Adaboost algorithm. In addition, it speeds up strong classifier converging and reduces the number of weak classifiers. For frontal face detection, the experiments on MIT+CMU frontal face test set result a 96.4% correct rate with 528 false alarms; for multi-view face in real time test set result a 94.7 % correct rate. The experimental results verified the effectiveness of the proposed approach.