Search

Character segmentation from License plate is challenging as it suffers from non-uniform illumination, blur and touching adjacent character due to head light of vehicles effect etc. This paper presents a new approach based on sharpness of character and the space regions for segmentation. We explore Gradient Vector Flow Opposite Direction Pair of pixels for seed point's selection. From the seed points, the proposed approach detects probable cuts between character components based on minimum cost path estimation. Then we propose a new sharpness for detecting candidate cuts from the probable cuts based on Laplacian zero crossing points and Sobel-Gradient. The average width of candidate cuts is considered as reference cut to identify the correct cuts between character components. Experimental results on real industry dataset of license plate shows that the proposed approach is robust to touching, blur, poor quality images. Further, comparative study with an existing approach shows that the proposed approach outperforms the existing approach.

This paper introduces a new method to estimate the instantaneous heart rates from facial images obtained from video camera. In this study, facial images of a subject whose heart rate varying from 150 beats per minute (BPM) to 110 BPM in 100 seconds are captured through video camera and analyzed using filter bank. The filter bank is utilized to provide the localized temporal information. We test the proposed method on videos with duration of 20s. The proposed method is compared to short-time Fourier Transform, which can also give temporal information. Experimental results show the proposed method can provide an acceptable result. The root mean square error is about 2.31 beats per minute (BPM). The average percentage error is 1.42% while the correlation coefficient 0.96.

This paper presents an image block classification method using Tchebichef moments (TMs) and support vector machine (SVM). The test images are divided into non-overlapping 16 × 16 blocks and transformed into moment domain using Discrete Tchebichef Transform. These moment features are then used in the image content (block) classification. SVM is used for learning and classifying the blocks into three types: “plain”, “edge” and “texture”, based on their moment energy level. Experimental results show that the proposed method works well and the classification accuracy is 98.7%.

Non-invasive heart rate measurement is essential in medical and sport sciences. Typically, the first stage of heart rate measurements is obtained from non-invasive sensors. In this paper, the data from video sequences of a subject cycling are processed using short-time Fourier transform (STFT) to indicate the heart rate of the subject. The choice of using STFT is its ability to provide more accurately localized temporal and frequency information, especially for the rapidly changing heart rate pattern during the exercise routine. Experimental results show the proposed method can provide an acceptable result and the root mean square error is less than 2.5 beats per minute (BPM) for heart rate variations between 80 BPM and 130 BPM.

In this paper, a face recognition system based on Discrete Tchebichef-Krawtchouk Transform DTKT and Support Vector Machines SVMs is proposed. The objective of this paper is to present the following: (1) the mathematical and theoretical frameworks for the definition of the DTKT including transform equations that need to be addressed. (2) the DTKT features used in the classification of faces. (3) results of empirical tests that compare the representational capabilities of this transform with other types of discrete transforms such as Discrete Tchebichef transform DTT, discrete Krawtchouk Transform DKT, and Discrete Cosine transform DCT. The system is tested on a large number of faces collected from ORL and Yale face databases. Empirical results show that the proposed transform gives very good overall accuracy under clean and noisy conditions.

On behalf of the technical program committee, we are pleased and honored to invite and welcome you to Malaysia and to the 2010 IEEE Asia Pacific Conference on Circuits and Systems (APCCAS 2010). In APCCAS 2010, the technical program shows the novel and advanced development in the field of circuits and systems. We also hope all attendees have fun in enjoying the history and cultures of Malaysia.

Image quality assessment is one of the challenging field of digital image processing system. It can be done subjectively or objectively. PSNR is the most popular and widely used objective image quality metric but it is not correlate well with the subjective assessment. Thus, there are a lot of objective image quality metrics (IQM) developed in the past few decades to replace PSNR. This paper provides a literature review of the current subjective and objective image quality measures. The purpose of this paper is to collect reported quality metrics and group them according to their strategies and techniques.

A non-invasive blood glucose monitoring system with six laser diodes is used to obtain a total of 290 NIR spectra from the Oral Glucose Tolerance Test (OGTT) experiment with the participation of a healthy volunteer over 4 days. Each laser diode operates at the discrete wavelengths between 1500nm and 1800nm with the power of 6mW each. A comparative analysis using the Partial Least Squares (PLS) model and the Neural Network (NN) model is studied. The study shows that the NN model performs better than the PLS model due to the presence of nonlinearity in the collected data. The presence of the nonlinearity is tested by using the Durbin-Watson test.

A common approach in mapping a signal to discrete events is to define a set of symbols that correspond to useful acoustic features of the signal over a short constant time interval. This paper proposes a Hidden Markov Models (HMM) based speech recognition by using cepstrum feature of the signal over adaptive time interval. First pitch period is detected by dyadic wavelet transform and divides the voiced speech signal according to the detected period. Then, system performs HMM-based speech recognition using cepstrum feature to classify the speech signals. Two speech recognition systems have been developed, one is based on constant time framing and the other is adaptive framing. The results are compared and found that adaptive framing method shows better result in both data distribution and recognition rate.

This paper presents the face recognition using discrete orthogonal moment that is Krawtchouk moments (KMs) on combined features of global and local face images. KMs are considered due to their ability to localize face image according to Region of Interest (ROI) unlike other moments which generally capture the global features. To obtain the global face image, both parameters of KMs are set equal at 0.5 while the parameters for local face image are set unequal. The selection of orders and parameters of KMs determines the ROI. The experiments carried out utilized the database face images from Olivetti research laboratory (ORL) consisting of 40 subjects of 10 non-similar images each [1]. Each face image differs in terms of position, rotation, scale and expression, with and without glasses. Euclidean square distance or Nearest Neighbour (NN) is used as the classifier in the recognition stage. From the experiments, the combined features of global and local faces images using KMs showed improved classification accuracy.

Changes in illumination condition, pose, facial expression and others is not an easy task in face recognition. Solving these problems requires not only a feature extraction method that can generate distinct features for each class of image but requires other additional technique that able to improve the overall classification accuracy. This paper presents the face recognition using combined features of original and pre-processed face images. This technique is experimented using orthogonal moments namely Zernike moments (ZMs) and Krawtchouk moments (KMs). The classification technique used in the recognition stage is Euclidean square distance or Nearest Neighbour (NN) classifier. Database face images from Olivetti research laboratory (ORL) consisting of 40 subjects of 10 images each where none of them are identical, is used in the experiments. The face images vary in position, rotation, scale and expression, with and without spectacles. From the experiments, the new technique is able to improve the classification accuracy significantly.

H.264/AVC exploits temporal and spatial redundancies to obtain higher coding gain. This comes with a cost of higher complexity encoder and higher computational time, which makes it difficult for real-time applications. In this paper, we present a fast intra-prediction mode selection method for H.264/AVC based on local edge information. Prior to intra prediction process, edge vectors of every pixel in a 4x4 block are totaled up to determine the edge direction. Based on the edge direction information, one directional mode is chosen and compared with the DC mode to choose the best mode with least rate distortion cost. Experimental results show that the proposed method reduces the computation complexity while it gives an almost similar visual quality of the reconstructed frame as compared to the one obtained by using exhaustive prediction method. The proposed method also outperforms Feng Pan's method (2005) in computational time and number of bits generated

Recognizing face images due to changes in illumination condition, pose, facial expression and others are challenging task. Solving these problems requires a feature extraction method that can generate distinct features for each class of image. Hence, this paper describes the comparative analysis of feature extraction methods namely geometric moments, Zernike moments, Krawtchouk moments and principle component analysis (PCA) in terms of their capability to recognize face images. The classification technique employed in the recognition stage is back propagation neural network (BPNN). The experiments utilized database face images from Olivetti research laboratory (ORL) consisting of 40 subjects of 10 samples each where none of them are identical. They vary in position, rotation, scale and expression. From the comparative study, the most suitable feature extraction method is considered for face recognition system.

This paper presents an algorithm in which the moment invariant features of a numerical character are presented as inputs to a neural network. The image is digitised, by a frame grabber card plugged into the extension slot of an IBM AT computer. Three moment invariant features derived from the geometrical moments were extracted and fed to the neural network. The neural network, used back propagation learning in its training, classifies these features and attempts to produce the desired output. Noisefree and noisy images of various variances of the ten numerals were experimented and the results are reported.

Genetic algorithm is used to perform variable selection to determine the ranges of wavelengths in NIR spectral data suitable to be used as predictors in multivariate calibration model via partial least squares. The NIR spectral data consists of three components of active substances, namely human serum albumin (HSA),#947;-globulin and glucose. The wavelength selection is able to improve the calibration model by selecting the wavelengths that contains information or correlated with the concentration of substances, while others non-chosen wavelengths, which contribute no information or contain noises, are excluded from the calibration model.

Human crowds organize themselves spontaneously in a confined environment when crowd density exceeds certain critical value. This self-organization behaviour of crowd motion can be observed from formation of walking lane with similar walking direction. In this work, we proposed a modified social force model, which is originally proposed by Helbing. Individuality, irregularity and unpredictability are characteristics of crowd motion. These properties are chaotic properties of nonlinear dynamical equation. Using this equation, chaotic perturbation is introduced into social force model to take into account the properties of crowd motion. It is shown that individuality, irregularity and unpredictability of crowd motion can eventually lead the crowd into self-organization.

Proposes a technique that uses genetic algorithm (GA) to select optimal features for classification applications using fuzzy ARTMAP (FA) neural network (NN). The technique is applied to select features for classification of two groups of subjects: alcoholics and controls, using multi-channel single trial electroencephalogram (EEG) signals evoked during visual response. The results show that the proposed technique is successful in selecting the features that contribute towards classification. This serves to reduce the number of required features while improving the classification performance. The results also indicate that the gamma band spectral power could be used to support evidence on the residual effects of long-term use of alcohol on visual response.

Presents a technique to classify signatures produced by the pressure exerted on the pen tip. Before the features are extracted, a low-pass filter is designed to remove frequencies greater that 50 Hz. A segmentation method using moving-average filtering and gradient calculation is used to divide the time series data into segments. The autoregressive (AR) coefficients are derived from each segment. From the coefficients, the power spectral density (PSD) is determined for every segment. A genetic algorithm (GA) is used to select the range of frequencies that contains the most important information. The selected range of frequencies is then fed into a multilayer perceptron (MLP) classifier with one hidden layer for verification. A database of 1,000 signatures is used for training and testing. The system is tested for genuine as well as forged signatures. The result obtained showed an average error rate of 3.09-3.33%.

Krawtchouk polynomials can be used to form the basis of a set of discrete orthogonal moments. In this paper, Krawtchouk moments are used to analytically reconstruct binary images. Their performance is compared to that of Zernike, Legendre (continuous) and Chebyshev (discrete) moments.

It is shown that by utilizing the method of Tchebichef (also know as Chebyshev) moments, an image degraded by Gaussian, salt-and-pepper and speckle noise can be restored. The performance of Tchebichef moments is compared with that of Zernike and Legendre moments.

Regular moment invariant face two limitations. First, images with symmetry in the x and/or y directions and symmetry at centroid give zero values for odd orders of central moments. Secondly, they are very sensitive to noise, especially the higher order moments. This paper presents a single solution to solve the symmetrical problem and reduce the noise sensitivity of these moments. The solution involves a new set of moment-based features that uses a reference point other than the image centroid. The reference centre is selected such that the new moment features are invariant to translation, scaling and rotation. The derivation of the new moments and their invariance are shown before experimenting them with some symmetrical alphabets. Next, they are shown to be less sensitive under the presence of Gaussian and random noise as compared to the usual regular moment invariants. Noise corrupted English alphabets are classified with a neural network to further verify the advantage of using the new moment features.

Automatic focusing of a visual inspection system is described. A vidicon camera was mounted on top of a high powered microscope to capture an image, process it and send an appropriate signal to a stepper motor which controls the movements of the microscope. The method used is based on the acutance of the video signal waveform. Acutance is a measure of the steepness of the video signal waveform. It implies that a sharp image will have a sharp gradient of the gray level across the image. This is measured vertically and horizontally and used to control the movements of a stepper motor to achieve an optimal focusing state. On achieving it a signal from the control circuitry of the automatic focusing system will be sent to the control circuitry of the die inspection machine to determine whether the die is good or a reject. >

This paper presents a new technique to derive features for images that are translated, scaled equally/unequally and rotated. The problem is formulated using conventional regular moments. It is shown that the conventional regular moment-invariants remain no longer invariant when the image is scaled unequally in the x- and y-directions. A method is proposed to form moment-invariants that do not change under such unequal scaling. The newly formed moments are also invariant to translation and reflection. However, it is not invariant for images that are rotated. A neural network is trained to estimate the angle of rotation; it is then used to derive the invariant moments for images that are unequally scaled, translated and rotated. Computer simulation results are also included to show the validity of the method proposed.

This paper presents a technique to derive features for images that are shifted and unequally scaled. The problem is formulated using conventional regular moments. It is shown that the conventional regular moment-invariants remain no longer invariant when the image is scaled unequally in the x- and y-directions. A method is proposed to form moment-invariants that do not change under such unequal scaling. Computer simulation results are also included to show the validity of the method proposed.

A new method of communication is proposed for paralysed patients using EEG signals. A scheme based on Morse code is used to construct meaningful words using recognised mental task. As the technology advances, it is envisaged that this technique could be used by anyone for rudimentary user-interface actions, like popping up windows and making menu choices. The EEG signals are segmented and power spectral values are extracted using Wiener-Khintchine theorem with Parzen window smoothing. A Fuzzy ARTMAP (FA) classifier is used to classify these signals into three different mental tasks, where each task either represents dit, dah or space. We analysed different mental task and show that the performance with different task varies greatly for different subjects. The FA classification results obtained show that it is highly possible to recognise three different mental task for each subject provided that these three tasks, which varies for different subjects are chosen after some preliminary simulations.

We describe a method to extract single trial visual evoked potential (VEP) buried in ongoing electroencephalogram (EEG) activity. The common method for separating VEP from EEG is to use signal averaging. But we use digital filters to extract VEP assuming that VEP spectra are in the gamma band. As an application, a fuzzy ARTMAP (FA) neural network classifier with voting strategy is used with this extracted VEP to discriminate alcoholics from normal subjects. The VEP is extracted from subjects while seeing visuals of the Snodgrass and Vanderwart (1980) picture set. The high FA classification of 96.5% shows the validity of the proposed method to successfully remove EEG contamination.

Visual inspection of EEG signals in their unprocessed form is still the predominant way of discriminating EEG patterns in the medical community and requires highly trained medical professionals. To overcome this problem, automatic EEG analysis using Fourier transform methods are popular since most EEG signals consist of the spectral power in the range of /spl delta/, /spl theta/, /spl alpha/ and /spl beta/, i.e. from 0 to 30 Hz. However, this method suffers from high noise sensitivity and is not suitable for short and variable length signal segments. We analyse EEG signals with time series analysis using autoregression techniques. We classify these extracted features for different mental tasks using a fuzzy ARTMAP classifier. We study the effects of different EEG segment or window lengths and different overlapping lengths on the overall performance of the classifier. Our results show that the segment length affects the performance and that overlapping the segments improves the performance greatly.

This paper presents a technique to classify images that have been elongated or contracted. The problem is formulated using conventional regular moments. It is shown that the conventional regular moment-invariants remain no longer invariant when the image is scaled unequally in the x- and y-directions. A method is proposed to form moment-invariants that do not change under such unequal scaling. Results of computer simulations for images are also included verifying the validity of the method proposed. >