• The Journal of the Korea institute of electronic communication sciences
• /
• v.8 no.8
• /
• pp.1129-1135
• /
• 2013

A new vehicle detection algorithm based on the multiple optimal Adaboost classifiers with optimal feature selection is proposed. It consists of two major modules: 1) Theoretical DDISF(Distance Dependent Image Scaling Factor) based image scaling by site modeling of the installed cameras. and 2) optimal features selection by Haar-like feature analysis depending on the distance of the vehicles. The experimental results of the proposed algorithm shows improved recognition rate compare to the previous methods for vehicles and non-vehicles. The proposed algorithm shows about 96.43% detection rate and about 3.77% false alarm rate. These are 3.69% and 1.28% improvement compared to the standard Adaboost algorithmt.

• Journal of Korea Multimedia Society
• /
• v.18 no.11
• /
• pp.1289-1301
• /
• 2015

Pedestrian detection (PD) is an essential task in various applications and sliding window-based methods utilizing HOG (Histogram of Oriented Gradients) or HOG-like descriptors have been shown to be very effective for accurate PD. However, due to exhaustive search across images, PD methods based on sliding window usually require heavy computational time. In this paper, we propose a real-time PD method for embedded visual surveillance with fixed backgrounds. The proposed PD method employs HOG descriptors as many PD methods does, but utilizes selective search so that it can save processing time significantly. The proposed selective search is guided by restricting searching to candidate regions extracted from Adaptive Gaussian Mixture Model (AGMM)-based background subtraction technique. Moreover, approximate computation of HOG descriptor and implementation in fixed-point arithmetic mode contributes to reduction of processing time further. Possible accuracy degradation due to approximate computation is compensated by applying an appropriate one among three offline trained SVM classifiers according to sizes of candidate regions. The experimental results show that the proposed PD method significantly improves processing speed without noticeable accuracy degradation compared to the original HOG-based PD and HOG with cascade SVM so that it is a suitable real-time PD implementation for embedded surveillance systems.

• Proceedings of the Korea Inteligent Information System Society Conference
• /
• 2007.05a
• /
• pp.309-320
• /
• 2007

The data imbalance problem which can be uncounted in data mining classification problems typically means that there are more or less instances in a class than those in other classes. It causes low prediction accuracy of the minority class because classifiers tend to assign instances to major classes and ignore the minor class to reduce overall misclassification rate. In order to solve the data imbalance problem, there has been proposed a number of techniques based on resampling with replacement, adjusting decision thresholds, and adjusting the cost of the different classes. In this paper, we study the feasibility of the combination usage of the techniques previously proposed to deal with the data imbalance problem, and suggest a combination method using genetic algorithm to find the optimal combination ratio of the techniques. To improve the prediction accuracy of a minority class, we determine the combination ratio based on the F-value of the minority class as the fitness function of genetic algorithm. To compare the performance with those of single techniques and the matrix-style combination of random percentage, we performed experiments using four public datasets which has been generally used to compare the performance of methods for the data imbalance problem. From the results of experiments, we can find the usefulness of the proposed method.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.15 no.6
• /
• pp.1213-1220
• /
• 2011

In this paper, a new ensemble system using dynamic weighting method with added weight information into classifiers is proposed. The weights used in the traditional ensemble system are those after the training phase. Once extracted, the weights in the traditional ensemble system remain fixed regardless of the test data set. One way to circumvent this problem in the gating networks is to update the weights dynamically by adding processes making architectural hierarchies, but it has the drawback of added processes. A simple method to update weights dynamically, without added processes, is proposed, which can be applied to the already established ensemble system without much of the architectural modification. Experiment shows that this method performs better than AdaBoost.

• The KIPS Transactions:PartD
• /
• v.9D no.5
• /
• pp.939-944
• /
• 2002

In this paper, we propose and apply new classification method to the remotely sensed image acquired from airborne multi-spectral scanner. This is a neuro-fuzzy image classifier derived from the generic model of a 3-layer fuzzy perceptron. We implement a classification software system with the proposed method for land cover image classification. Comparisons with the proposed and maximum-likelihood classifiers are also presented. The results show that the neuro-fuzzy classification method classifies more accurately than the maximum likelihood method. In comparing the maximum-likelihood classification map with the neuro-fuzzy classification map, it is apparent that there is more different as amount as 7.96% in the overall accuracy. Most of the differences are in the "Building" and "Pine tree", for which the neuro-fuzzy classifier was considerably more accurate. However, the "Bare soil" is classified more correctly with the maximum-likelihood classifier rather than the neuro-fuzzy classifier.

• Journal of Information Processing Systems
• /
• v.9 no.1
• /
• pp.173-188
• /
• 2013

In Facial Expression Recognition Systems (FERS), only particular regions of the face are utilized for discrimination. The areas of the eyes, eyebrows, nose, and mouth are the most important features in any FERS. Applying facial features descriptors such as the local binary pattern (LBP) on such areas results in an effective and efficient FERS. In this paper, we propose an automatic facial expression recognition system. Unlike other systems, it detects and extracts the informative and discriminant regions of the face (i.e., eyes, nose, and mouth areas) using Haar-feature based cascade classifiers and these region-based features are stored into separate image files as a preprocessing step. Then, LBP is applied to these image files for facial texture representation and a feature-vector per subject is obtained by concatenating the resulting LBP histograms of the decomposed region-based features. The one-vs.-rest SVM, which is a popular multi-classification method, is employed with the Radial Basis Function (RBF) for facial expression classification. Experimental results show that this approach yields good performance for both frontal and near-frontal facial images in terms of accuracy and time complexity. Cohn-Kanade and JAFFE, which are benchmark facial expression datasets, are used to evaluate this approach.

• Journal of KIISE:Software and Applications
• /
• v.31 no.8
• /
• pp.999-1009
• /
• 2004

As a new approach to the diagnosis of cancers, bioinformatics attracts great interest these days. Machine teaming techniques have produced valuable results, but the field of medicine requires not only highly accurate classifiers but also the effective analysis and interpretation of them. Since gene expression data in bioinformatics consist of tens of thousands of features, it is nearly impossible to represent their relations directly. In this paper, we propose a method composed of a feature selection method and genetic programming. Rank-based feature selection is adopted to select useful features and genetic programming based arithmetic operators is used to generate classification rules with features selected. Experimental results on Lymphoma cancer dataset, in which the proposed method obtained 96.6% test accuracy as well as useful classification rules, have shown the validity of the proposed method.

• Proceedings of the Korean Society for Bioinformatics Conference
• /
• 2003.10a
• /
• pp.133-138
• /
• 2003

The prediction of protein secondary structure has been an important bioinformatics tool that is an essential component of the template-based protein tertiary structure prediction process. It has been known that the predicted secondary structure information improves both the fold recognition performance and the alignment accuracy. In this paper, we describe several novel ideas that may improve the prediction accuracy. The main idea is motivated by an observation that the protein's structural information, especially when it is combined with the evolutionary information, significantly improves the accuracy of the predicted tertiary structure. From the non-redundant set of protein structures, we derive the 'potential' parameters for the protein secondary structure prediction that contains the structural information of proteins, by following the procedure similar to the way to derive the directional information table of GOR method. Those potential parameters are combined with the frequency matrices obtained by running PSI-BLAST to construct the feature vectors that are used to train the support vector machines (SVM) to build the secondary structure classifiers. Moreover, the problem of huge model file size, which is one of the known shortcomings of SVM, is partially overcome by reducing the size of training data by filtering out the redundancy not only at the protein level but also at the feature vector level. A preliminary result measured by the average three-state prediction accuracy is encouraging.

• Asia pacific journal of information systems
• /
• v.25 no.3
• /
• pp.579-596
• /
• 2015

Many studies have focused on the prediction of corporate credit rating using various data mining techniques. One of the most frequently used algorithms is support vector machines (SVM), and recently, novel techniques such as SVM+ and SVM+MTL have emerged. This paper intends to show the applicability of such new techniques to multi-classification and corporate credit rating and compare them with conventional SVM regarding prediction performance. We solve multi-class SVM+ and SVM+MTL problems by constructing several binary classifiers. Furthermore, to demonstrate the robustness and outstanding performance of SVM+MTL algorithm over other techniques, we utilized four typical multi-class processing methods in our experiments. The results show that SVM+MTL outperforms both conventional SVM and novel SVM+ in predicting corporate credit rating. This study contributes to the literature by showing the applicability of new techniques such as SVM+ and SVM+MTL and the outperformance of SVM+MTL over conventional techniques. Thus, this study enriches solving techniques for addressing multi-class problems such as corporate credit rating prediction.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.11
• /
• pp.5594-5615
• /
• 2019

Malicious social robots, which are disseminators of malicious information on social networks, seriously affect information security and network environments. The detection of malicious social robots is a hot topic and a significant concern for researchers. A method based on classification has been widely used for social robot detection. However, this method of classification is limited by an unbalanced data set in which legitimate, negative samples outnumber malicious robots (positive samples), which leads to unsatisfactory detection results. This paper proposes the use of generative adversarial networks (GANs) to extend the unbalanced data sets before training classifiers to improve the detection of social robots. Five popular oversampling algorithms were compared in the experiments, and the effects of imbalance degree and the expansion ratio of the original data on oversampling were studied. The experimental results showed that the proposed method achieved better detection performance compared with other algorithms in terms of the F1 measure. The GAN method also performed well when the imbalance degree was smaller than 15%.