Tae-Seung Lee

Improved edge enhanced error diffusion based on first-order gradient shaping filter

One of the difficulties of using Artificial Neural Networks (ANNs) to estimate atmospheric temperature is the large number of potential input variables available. In this study, four different feature extraction methods were used to reduce the input vector to train four networks to estimate temperature at different atmospheric levels. The four techniques used were: genetic algorithms (GA), coefficient of determination (CoD), mutual information (MI) and simple neural analysis (SNA). The results demonstrate that of the four methods used for this data set, mutual information and simple neural analysis can generate networks that have a smaller input parameter set, while still maintaining a high degree of accuracy.

Distortion-Free of General Information with Edge Enhanced Error Diffusion Halftoning

The error diffusion method is good for reconstructing continuous tones of an image to bilevel tones. However, the reconstruction of edge information by the error diffusion is represented as weak when the power spectrum is analyzed for display error. In this paper, we present an edge enhanced error diffusion method to preprocess original images to achieve an enhancement for the edge information. The preprocessing algorithm consists of two processes. First, the value of difference between the current pixel and the local average of surrounding pixels in the original image is obtained. Second, weighting function is composed by the magnitude and the sign of the local average. To confirm the effect of proposed method, the method is compared with the standard error diffusion and conventional edge enhanced error diffusion methods by measuring various objective measuring criteria including the radially averaged power spectrum density (RAPSD) for display error. The results of comparison demonstrate the superiority of the proposed method over the conventional ones.

Continuants based neural speaker verification system

Among the techniques to protect private information by adopting biometrics, speaker verification is widely used due to its advantages in natural usage and inexpensive implementation cost. Speaker verification should achieve a high degree of reliability in verification score, flexibility in speech text usage, and efficiency in the complexity of verification system Continuants have an excellent speaker-discriminant power and the modest number of phonemes in the phonemic category. Multilayer perceptrons (MLPs) have the superior recognition ability and the fast operation speed. In consequence, the two elements can provide viable ways for speaker verification system to obtain the above properties: reliability, flexibility and efficiency. This paper shows the implementation of a system to which continuants and MLPs are applied, and evaluates the system using a Korean speech database. The results of the evaluation prove that continuants and MLPs enable the system to acquire the three properties.

Implementation of traffic flow measuring algorithm using real-time dynamic image processing

In this paper, an automatic traffic flow measuring system and a real-time image processing algorithm have been developed. The picture of moving vehicles taken by an industrial television (ITV) camera are digitized into sample points in odd ITV frames and the points are processed in even ITV frames by a personal computer. We detect the presence of vehicles by comparing the brightness of the sample points of vehicle with that of the road. After eliminating noises contained in the digitized sample points by appropriate smoothing techniques, we obtain a contour of each vehicle. Using the contour, the number of passing vehicles is effectively measured by counting the number of sample points of each vehicle. Also the type of a vehicle is easily figured out by counting the number of sample points corresponding to the width of vehicles contour. The performance of the proposed algorithm is demonstrated by actual implementation. From the experimental results 1-2% measurement error was observed.

A Qualitative Discriminative Cohort Speakers Method to Reduce Learning Data for MLP-Based Speaker Verification Systems

Although multilayer perceptrons (MLPs) present several advantages against other pattern recognition methods, MLP-based speaker verification systems suffer from slow enrollment speed caused by many background speakers to achieve a low verification error. To solve this problem, the quantitative discriminative cohort speakers (QnDCS) method, by introducing the cohort speakers method into the systems, reduced the number of background speakers required to enroll speakers. Although the QnDCS achieved the goal to some extent, the improvement rate for the enrolling speed was still unsatisfactory. To improve the enrolling speed in this paper, the qualitative DCS (QlDCS) has been proposed by introducing a qualitative criterion to select less background speakers. An experiment for both methods is conducted to use the speaker verification system based on MLPs and continuants, and speech database. The results of the experiment show that the proposed QlDCS method enrolls speakers in shorter time than the QnDCS does.

Faster speaker enrollment for speaker verification systems based on MLPs by using discriminative cohort speakers method

Speaker verification system has been currently recognized as an efficient security facility due to its cheapness and convenient usability. This system has to achieve fast enrollment and verification in order to make a willing acceptance to users, as well as low error rate. For accomplishing such low error rate, multilayer perceptrons (MLPs) are expected to be a good recognition method among various pattern recognition methods for speaker verification. MLPs process speaker verifications in modest speed even with a low-capable hardware because they share their internal weights between all recognizing models. On the other hand, considerable speaker enrolling delay is made mainly due to the large population of background speakers for low verification error, since the increasing number of the background speakers prolongs the learning times of MLPs. To solve this problem, this paper proposes an approach to reduce the number of background speakers needed to learn MLPs by selecting only the back ground speakers nearby to an enrolling speaker. An experiment is conducted using an MLP-based speaker verification system and Korean speech database. The result of the experiment shows efficient improvement of 23.5% in speaker enrolling time.

Toward more reliable emotion recognition of vocal sentences by emphasizing information of korean ending boundary tones

Autonomic machines interacting with human should have capability to perceive the states of emotion and attitude through implicit messages for obtaining voluntary cooperation from their clients. Voice is the easiest and the most natural way to exchange human messages. The automatic systems capable of understanding the states of emotion and attitude have utilized features based on pitch and energy of uttered sentences. Performance of the existing emotion recognition systems can be further improved with the support of linguistic knowledge that specific tonal section in a sentence is related to the states of emotion and attitude. In this paper, we attempt to improve the recognition rate of emotion by adopting such linguistic knowledge for Korean ending boundary tones into an automatic system implemented using pitch-related features and multilayer perceptrons. From the results of an experiment over a Korean emotional speech database, a substantial improvement is confirmed.

Elastic Learning Rate on Error Backpropagation of Online Update

The error-backpropagation (EBP) algorithm for learning multilayer perceptrons (MLPs) is known to have good features of robustness and economical efficiency. However, the algorithm has difficulty in selecting an optimal constant learning rate and thus results in non-optimal learning speed and inflexible operation for working data. This paper introduces an elastic learning rate that guarantees convergence of learning and its local realization by online update of MLP parameters into the original EBP algorithm in order to complement the non-optimality. The results of experiments on a speaker verification system with Korean speech database are presented and discussed to demonstrate the performance improvement of the proposed method in terms of learning speed and flexibility for working data of the original EBP algorithm.

Optimization of Usability on an Authentication System Built from Voice and Neural Networks

While multilayer perceptrons (MLPs) have great possibility on the application to speaker verification, they suffer from an inferior learning speed. To appeal to users, the speaker verification systems based on MLPs must achieve a reasonable speed of user enrolling and it is thoroughly dependent on fast learning of MLPs. To attain real-time enrollment for the systems, the previous two studies, the discriminative cohort speakers (DCS) method and the omitting patterns in instant learning (OIL) method, have been devoted to the problem and each satisfied that objective. In this paper, we combine the two methods and apply the combination to the systems, assuming that the two methods operate on different optimization principles. Through experiment on real speech database using an MLP-based speaker verification system to which the combination is applied, the feasibility of the combination is verified from the results. Keywords: Biometric authentication system, speaker verification, multiplayer perceptrons, error backpropagation, real-time enrollment, discriminative cohort speakers, omitting patterns in instant learning

Source estimating anycast for high quality of service of multimedia traffic

One of the difficulties of using Artificial Neural Networks (ANNs) to estimate atmospheric temperature is the large number of potential input variables available. In this study, four different feature extraction methods were used to reduce the input vector to train four networks to estimate temperature at different atmospheric levels. The four techniques used were: genetic algorithms (GA), coefficient of determination (CoD), mutual information (MI) and simple neural analysis (SNA). The results demonstrate that of the four methods used for this data set, mutual information and simple neural analysis can generate networks that have a smaller input parameter set, while still maintaining a high degree of accuracy.