Earticle

On speech signal processing, it is very important to find the fundamental frequency of voice. The reason is why it is used in variable places, such as speech-enhancement-system, speech-recognition system, speaker-classification-system, and handicapped assisting-system. However the pitch detection is difficult when the original signal is corrupted by noise, or put in transition section of voice. In this paper, we make proposal of the method that enhance accuracy of pitch detection system, through SNR compensation using time-domain SNR estimator with continuous voice signal. And we proved the performance of the detector, in drawing pitch contour of variable SNR signals.

There are various methods to improve signal under the clipped environment of voiced sound in the voice signal processing. In the case of the majority of clipping signal restoration, it will be concentrated in sound quality deterioration of low frequency as they are converged into voiced sound from the characteristic viewpoint of sound quality deterioration. Therefore, in the case of clipping of voiced sound section, great efforts have been contributed for sound quality improvement and restoration. From this study, the method to improve clipped voice signal will be proposed. Voiced sound section is created by utilizing the fact that differentiation and clipped signal in voice signal are supposed to give small influence for the evaluation of recognition level. In here, for the signal which is obtained by stressing amplitude at voiced sound section where most of them are distributed, the method of emphasizing synthesis with original signal was used.

Robust detection of humans in image sequences is important for many applications. However, if humans are adjacent to each other, it is much more difficult to accurately detect them. In this paper, we propose a method to automatically detect multiple humans using motion information and Adaboost algorithm from a single camera on a mobile or stationary system. In case of mobile system, the ego-motion of the camera is compensated by the corresponding feature sets. The region of interest that moving objects are likely to exist is searched by the projection approach using a difference image between two consecutive images that an ego-motion is compensated. Human detector is learned by boosting a number of weak classifiers which are based on Harr-like features. The proposed approach has been tested to a number of image sequences, and it was shown to detect multiple humans very well.

This paper proposes a new framework for incremental learning based on accuracy and coverage. Classied addition of example into four cases, two inequalities for accuracy and coverage are obtained. The proposed method classies a set of formulae into three layers: rule layer, subrule layer and non-rule layer by using the inequalities obtained. Then, subrule layer plays a central role in updating rules.

Artificial neural network (ANN) have wide applicability in various applications in the life, one of them is apply to recognize characters and numbers, and we know that the Automatic License Plate Recognition (ALPR) is very important in the Intelligent Transportation System (ITS) and it is beginning in research and application in Vietnam. Usually, an ALPR system consists of three parts: 1) license plate location, 2) character segmentation, 3) characters & numbers recognition. In this paper, we proposed an improved method for the characters & numbers recognition part. And then, we apply to recognize characters & numbers of Vietnam License Plates (LP), which combined neural network and image processing technologies. In the training work, we used two networks and back-propagation (BP) algorithm for characters & numbers training with noises, separately, which the computing time and accuracy will be improved. In the using network work, we used the image processing technology for pre-processing to obtain high quality of characters & numbers before put in the trained network to improve accuracy of the system. We tested on 600 Vietnam LP images, which obtained from the actual systems, these images are very different background such as illumination, license angles, size and type, colors, light conditions in Vietnam environment. Our approach is more effective than some of the existing methods and satisfied for all types and color of Vietnam license plates and Vietnam environment.

In this paper, we present a smart grid monitoring system connected with electric vehicle charging system using anti-islanding method. Electric vehicles can be charged through remote control of smart grid monitoring system and the charging process may be more stable and more efficient by virtue of wireless communication between the Local Area Module and End Modules. It is illustrated by some experiments that electric vehicle charging process continued without any serious fault even though the islanding phenomena occurred in the grid if the presented monitoring system was applied to the smart grid system.

Recently, camera model identification becomes one of the most popular research topics in digital forensics field. Since every camera imaging processing left artifacts on its final output image, and some of them can be considered as model-specific ‘traces’ of its source camera, camera model can be classified only with a single image by catching these ‘traces’. This paper presents a camera model identification method based on multi-step transition matrices. We firstly model JPEG image coefficients by Markov process. Then, one-step and two-step transition matrices along different directions are extracted respectively. Finally, 58 statistics calculated from these matrices are used to perform camera model identification as features. In our experiment, we chose images from seven camera models in Dresden Image Database as our experiment samples. Experiments results show that the average detection accuracy of this method can reach to 99.27%. Compared with previous Markov method, our approach can perform better only using 58-D features.