Dong-Hun Seo

A new approach for classification of fingerprint image quality

Fingerprint identification is a technology which has been widely accepted for personal identification in many areas such as criminal investigation, access control, and Internet authentication due to its uniqueness. Most available systems for fingerprint identification use the minutiae matching for identification. The performance of minutiae extraction algorithms relies heavily on the quality of the fingerprint image because the minutiae-based approach is very sensitive to noise or image-quality degradation. Poor-quality images result in spurious and missing features which degrades the performance of the identification system. So if the quality of the image can be examined first, the image with very poor quality can be rejected. The useful information can be combined into the procedure of post-processing and matching scheme to improve the identification process. Therefore it is desirable to design a classification scheme which is able to examine the quality of a fingerprint image before it is processed by the fingerprint identification systems. Decision-tree is one of the well-known data mining methods that are used in classification problems because of its fast and effective features. In this paper a new classifier based on decision tree theory for classification of fingerprint image quality is proposed. This new classifier has many advantages in solving the fingerprint quality classification problem. It can generate the rules with almost all of the original information from the classifier even after all of the original fingerprint images are lost. It means it will be not necessary to save all the original fingerprint images because the classifier can combine the rules with the new coming data to build a more precise classifier with more information to classify the fingerprint image quality. And the new classifier can give weight to different fingerprint images if they come from the sensors with different importance. The advantages of the new classifier make the fingerprint image classification system very powerful and feasible. The proposed method has a very good performance which is proved by the experiments.

Rule refinement with extended data expression

The rule refinement problem has been known to be one of the most difficult and complex problems. This paper presents a systematic rule refinement method that deals with the old rule directly with the new data, for the first time. To be able to do the rule refinement, the data are represented in the extended data expression, where an event has its weight of importance. To show how this can be done systematically, a decision tree classifier is used for the rule refinement. The weights of the events of the former rule are adjusted according to the depth of the tree merged with the collected new data set to form the new rule. Experiment shows that this approach, with properly designing the weight assignment procedure, is promising to enhance the performance of the inference engine by generating a rule with higher accuracy than the one from new data set only.

A new classifier applied to biological early warning systems for toxicity detection

Biological early warning systems(BEWS) has been developed in recent years. BEWS detects toxicity by tracking the physiologic responses of the whole organisms. In the paper, we apply the classification technique to the biological early warning systems and propose a new BEWS which is meaningful to biological field. Meanwhile, how to select the features in such classification application is also a contribution of this paper. By using the fractal dimension theory, we define the input features which represent the organism characteristics in non-toxic or toxic environment. The experiment results show that the proposed new bio-monitoring system is effective for environmental toxicity detection.

FLDF Based Decision Tree using Extended Data Expression

We introduce a classification algorithm which can be applied to a problem with a data set included a missing variable. In this algorithm we use data expansion treating it with a weight value and the probability techniques. It is applied to extending a classifier which is considered the optimal projection plane based on Fishers formula. For doing this, we derive equations from the procedure to be applied to the data expansion. The result is compared to that of different measurements by choosing one variable in the data set and then modifying the rate of missing and non-missing values in this selected variable. The result of a data set with non-missing variable compares with that of C4.5 which is known as a knowledge acquisition tool in machine learning.

A Mean Field Annealing Algorithm for Fuzzy Clustering

In the classical clustering, an item must entirely belong to a cluster. Fuzzy clustering, however, describes more accurately the ambiguous type of structure in data. Fuzzy clustering is useful for partitioning a set of objects into a certain number of groups by assigning the membership probabilities to each object. In fuzzy clustering, the membership of each datum in each cluster is represented by the membership matrix. In the proposed method, the elements of membership matrix are updated in parallel until they reach one of the global optimal solutions. It differs from the traditional fuzzy clustering methods. In classical fuzzy clustering, the centroid vectors of the clusters in the space are calculated, and then the membership probability matrix is determined, and the process is repeated until the optimum solution is found. By contrast, the method proposed here perturbs the membership probability, and determines whether the the perturbed state should be accepted or not according to the changes of the energy. One Variable Stochastic Simulated Annealing(OVSSA), a continuous valued version of the Mean Field Annealing(MFA) algorithm which is known as a massively parallel algorithm, is employed as an optimization technique. The MFA combines characteristics of the simulated annealing and the neural network and exhibits the rapid convergence of the neural network while preserving the solution quality afforded by Stochastic Simulated Annealing(SSA).

Solving multi-sensor problem with a new approach

Smart environments is a technological concept that, according to Mark Weiser is ldquoa physical world that is richly and invisibly interwoven with sensors, actuators, displays, and computational elements, embedded seamlessly in the everyday objects of our lives, and connected through a continuous network.rdquo But sometimes the data gathered from the sensors is with different importance. It means some sensors are more reliable than others for some reasons. For example some sensors may be in relatively bad environments and some of the gathered data is destroyed or ruined. How to deal with the information gathered from different sensors efficiently is an important multi-sensors problem. The existence of multi-sensors problem will degrade the learning quality of classification models. And almost all of the existing classifier can not deal with this problem. So handling multi-sensors problem is important and necessary for building a high quality classification model and smart environments. In this paper a new classifier capable of dealing with this multi-sensors problem is proposed and it has a very good performance which is proved by experiments. This classifier can combine the information gathered from different sensors efficiently and in can add the new coming data to make a more efficient classifier even all of the original data is lost. Because of all the advantages it has, the new classifier is proposed sincerely to apply into smart environments.

A Classifier Capable of Rule Refinement

In ubiquitous environment, too much information is generated from a lot of sensors, and people want to obtain the appropriately classified information from the information. Decision tree algorithm like C4.5 is much useful in the field of data mining or machine learning system. Because this is fast and deduces good result on the problem of classification. This paper proposes three methods using decision tree for solving a classification problem. First, this paper suggest about the extended data expression. Second, a classifier, UChoo, based on the extended data expression is described. Third, this paper is to describe about rule generation. The rules expressed in the newly suggested method have almost the same information contents as compared with the original data set. The information is gotten from the sensors becomes large amount of data as the ubiquitous computation environment develops, therefore it is impossible to keep all information in memory. However, using suggested method, this problem is solved smoothly without losing almost the information.

A New Classification Application of Biological Early Warning Systems for Toxicity Detection

Biological early warning systems (BEWS) which detects toxicity by tracking the physiologic responses of the whole organisms has been developed in recent years. In the paper, we firstly apply the C4.5 decision tree Algorithm to the biological early warning systems and propose a new BEWS which can detect toxicity by classifying the fleas behavior data. The new system is meaningful to both biological field and data mining. The new classification application includes decision tree algorithm and feature selection. The new BEWS system consists of training and test process. The experiment results show that the proposed new bio-monitoring system is available for environmental toxicity detection.

A classifier capable of rule refinement

In ubiquitous environment, too much information is generated from a lot of sensors, and people want to obtain the appropriately classified information from the information. Decision tree algorithm like C4.5 is very useful in the field of data mining or machine learning system. Because this is fast and deduces good result on the problem of classification. This paper proposes three methods using decision tree for solving a classification problem. Firstly, this paper suggest about the extended data expression for explaining a classifier, Uchoo. Secondly, a classifier, UChoo, is described. Thirdly, this paper is to describe about rule generation. The rules expressed in the newly suggested method have almost the same information contents as compared with the original data set. The information is gotten from the sensors becomes large amount of data as the ubiquitous computation environment develops, therefore it is impossible to keep all information in memory. However, using suggested method, this problem is solved smoothly with having almost the information.

Visualizing a multi-dimensional data set in a lower dimensional space

This paper presents a method of visualizing a multi-dimensional data set into a lower dimensional space, especially into a two-dimensional space, so that people can intuitively conceive the relations or the distance between the entities of the data. Kullback-Leibler divergence is used as the measure to evaluate the distance between the vectors of the probability distribution. The measured distance values are used to find the corresponding coordinates of the entities in a lower dimensional space. Here, the one variable stochastic simulated annealing (OVSSA) is employed as the optimization technique. Experiments show that this is a plausible way of visualizing the multi-dimensional data, letting people see the relations among the entities intuitively.