Expert System for Speaker Identification Using Lip Features with PCA

— Neural networks are frequently used as a classifier for tasks in many classifications. However there are disadvantages in terms of amount of training data required, and length of training time. This paper, develop an intelligent diagnosis system for zinc oxide (ZnO) surge arrester fault classification. First the features were extracted from 600 ZnO surge arrester thermal images and leakage currents. Then these features were presented to several neural network architectures to investigate the most suitable network model for classifying the ZnO surge arrester fault condition effectively. Three classification models were used namely feed forward back propagation (FFBP), radial basis function (RBF) and learning vector quantization (LVQ) algorithm. The performance of the networks was compared based on resulted of misclassify and correct rate. The method was evaluated using 24 testing datasets. Comparison results showed that LVQ was the best training algorithm for the ZnO surge arrester fault classification compared to the others system. Also the LVQ is faster than FFBP and RBF.

Three 100-compound spectra libraries have been used to evaluate artificial neural network classifications of functional groups. A near-IR gas-phase library was used to compare neural network classifications with those obtained by two-dimensional principal component analysis (PCA) score plots and by the use of the Mahalanobis distance metric based on multidimensional (score) vectors. The neural network using a radial basis function algorithm was able to correctly classify all aromatic and nonaromatic samples in a test set of 40 samples from the 100-compound library; PCA score plots were successful in separating ∼92% of the 100-compound library into aromatic and nonaromatic classes, whereas the Mahalanobis distance metric could not separate the in-class vs out-of-class aromatics in the library. Using principal component scores as input to the neural network training with 40 randomly selected samples, validating with 20 randomly selected samples, and testing with 40 randomly selected samples were performed in less than 5 s and produced perfect classifications. The neural network algorithm incorporating the radial basis function was then used to compare the information available in a near-IR spectral library of condensed-phase molecules with spectra of identical (or very similar) compounds in a mid-IR library. Results with the radial basis function were very good for both libraries, with classifications >85% in all cases. The near-IR library produced better results for aromatics (95 vs 88%), identical or very similar results for OH's (98%), alkyls (>85%), and halogens (98%), and poorer results for carbonyls (85 vs 98%). Better mid-IR results for carbonyls were anticipated due to the sharp band for carbonyl-containing compounds in the fingerprint region; however, the improved results for aromatics in the near-IR were not anticipated.

Neural network based detection of hard exudates in retinal images

Three learning phases for radial-basis-function networks

The prognosis of childhood nephrotic syndrome directly hinges on the accurate prediction of negative conversion days (NCDs). Therefore, this paper designs a hybrid approach of principal component analysis (PCA) and backpropagation (BP)-adaptive boosting (AdaBoost) neural network (NN), and applies the method to predict the NCDs for children with nephrotic syndrome. Specifically, PCA method was employed for dimension reduction. Six principal components were extracted from multiple physiological features, and taken as input variables of three-tiered neural networks. The boosted predictor of BP-AdaBoost model, together with three predictors of BP NN, support vector machine (SVM) and radial basis function (RBF) NN, was trained for NCDs prediction. The experimental results show that the predictor of BP-AdaBoost NN achieves the mean absolute error of 0.2334, the mean relative error of 3.2789%, the SD of 4.6804 and the RSD of 50.8053% in NCDs prediction, and it outperforms other predictors of BP NN, SVM and RBF NN on both accuracy and precision. Furthermore, comparison experiments are conducted on PCA processed testing data and raw testing data for BP-AdaBoost NN and demonstrate the excellent effect of PCA. The hybrid approach of PCA and BP-AdaBoost NN is simple and reliable for NCDs prediction of childhood nephrotic syndrome, and it can help pediatricians prognose childhood nephrotic syndrome accurately and further provide patients with better care and treatment.

Classification of non-alcoholic beer based on aftertaste sensory evaluation by chemometric tools

Instantaneous learning is a desirable feature in neural networks. This type of learning enables the network to be trained very quickly, typically in just one or two passes of the training set as opposed to hundreds or even thousands of passes for networks trained by an iterative process, such as the error backpropagation (BP) algorithm. This dissertation first reviews several existing types of neural networks with instantaneous learning capability. It then proposes a new network called the Fuzzy Classification (FC) Neural Network that can be trained with two passes of the training samples. The first pass assigns the synaptic weights for the input and output layers. The second pass determines the radius of generalization r for each training sample. The network exhibits fuzziness in two regards: (1) by fuzzification of the location of each training vector in the input space; and (2) by assigning fuzzy memberships of output classes to new input vectors. The operation of the FC network is analyzed from different perspectives, namely, separability of patterns, curve fitting and kernel regression. It is shown that one mode of operation of the FC network can be made to behave like a Radial Basis Function (RBF) network by proper choice of membership function and network parameter. The performance of the FC network as a pattern classifier is also discussed in a statistical framework. The generalization performance of the FC network is compared against that of the CC4 Comer Classification Neural Network, the RBF network, and the Multilayer Perceptron network trained by the BP algorithm. Experiments involving a Henon map time series, a Mackey-Glass time series, and a spiral pattern are used in the comparison. The performance of the FC network is found to be comparable to that of RBF and BP networks, and better than that of the CC4 network. Its performance is also more scalable than those of CC4 and RBF networks. Further, the time taken to design an FC network for any given problem is much shorter compared to the other three networks.

A new method based on the integration of principal component analysis (PCA) and radial basic function (RBF) neural network is put forward for selecting the real estate project. Firstly, principal component analysis (PCA) is used to reduce the evaluation index dimensions. And then, radial basic function (RBF) neural network is used to evaluate the real estate projects. In order to grasp this method better, finally, the paper provides a case to demonstrate the application of this method in selecting the real estate project. The case has shown that the method applied to select the real estate project is feasible and reliable.

A learning vector quantization(LVQ) neural network classifier is for the first time applied for SAR data classification, of which the training termination rule is modified to make it have both the ability of classification and class signature analysis. A very high land cover classification accuracy is achieved. Especially under the condition that texture is considered, almost all roads can be classified correctly, which cannot be identified by BP MLP neural network and ML classifier.

This paper proposes a framework for developing a broad variety of soft clustering and learning vector quantization (LVQ) algorithms based on gradient descent minimization of a reformulation function. According to the proposed axiomatic approach to learning vector quantization, the development of specific algorithms reduces to the selection of a generator function. A linear generator function leads to the fuzzy c-means (FCM) and fuzzy LVQ (FLVQ) algorithms while an exponential generator function leads to entropy constrained fuzzy clustering (ECFC) and entropy constrained LVQ (ECLVQ) algorithms. The reformulation of clustering and LVQ algorithms is also extended to supervised learning models through an axiomatic approach proposed for reformulating radial basis function (RBF) neural networks. This approach results in a broad variety of admissible RBF models, while the form of the radial basis functions is determined by a generator function. This paper shows that gradient descent learning makes reformulated RBF neural networks an attractive alternative to conventional feed-forward neural networks.

We compared linear and neural network models for estimating human signal detection performance from event-related potentials (ERP) elicited by task-relevant stimuli. Data consisted of ERPs and performance measures from five trained operators who monitored a radar display and detected and classified visual symbols at three contrast levels. The performance measure (PF1) was a composite of accuracy, speed, and confidence of classification responses. The ERPs, which were elicited by the symbols, were represented in the interval 0-1500 ms post-stimulus at three midline electrodes (Fz, Cz, Pz) using either principal component analysis (PCA) factors or coefficients of autoregressive (AR) models. We constructed individual models of PF1 from both PCA and AR representations using either linear regression or radial basis function (RBF) networks. Applying the normalized mean square error of approximation as a criterion, we found that the PCA representation was superior to AR and that RBF networks estimated PF1 much more accurately than linear regression. This suggests that nonlinear methods combined with suitable ERP feature extraction can provide more accurate and reliable estimates of display-monitoring performance than linear models.

Radial Basis Function (RBF) metamodels have recently attracted increased interest due to their significant advantages over other types of non-parametric metamodels. However, because of the interpolation nature of the RBF mathematics, the accuracy of the model may dramatically deteriorate if the training data set used contains duplicate information, noise or outliers. Also constructing the metamodel may be time consuming whenever the training data sets are large or a high dimensional model is required. In this paper, we propose a robust and efficient RBF metamodeling approach based on data pre-processing techniques that alleviate the accuracy and efficiency issues commonly encountered when RBF models are used in typical real engineering situations. These techniques include 1) the removal of duplicate training data information, 2) the generation of smaller uniformly distributed subsets of training data from large data sets and 3) the quantification and identification of outliers by principal component analysis (PCA) and Hotelling statistics. Simulation results are used to validate the generalization accuracy and efficiency of the proposed approach.

Although the field of automatic speaker or speech recognition has been extensively studied over the past decades, the lack of robustness has remained a major challenge. The missing data technique (MDT) is a promising approach. However, its performance depends on the correlation across frequency bands. This paper presents a new reconstruction method for feature enhancement based on the trait. In this paper, the degree of concentration across frequency bands is measured with principal component analysis (PCA). Through theoretical analysis and experimental results, it is found that the correlation of the feature vector extracted from the sub-band (SB) is much stronger than the ones extracted from the full-band (FB). Thus, rather than dealing with the spectral features as a whole, this paper splits full-band into sub-bands and then individually reconstructs spectral features extracted from each SB based on MDT. At the end, those constructed features from all sub-bands will be recombined to yield the conventional mel-frequency cepstral coefficient (MFCC) for recognition experiments. The 2-sub-band reconstruction approach is evaluated in speaker recognition system. The results show that the proposed approach outperforms full-band reconstruction in terms of recognition performance in all noise conditions. Finally, we particularly discuss the optimal selection of frequency division ways for the recognition task. When FB is divided into much more sub-bands, some of the correlations across frequency channels are lost. Consequently, efficient division ways need to be investigated to perform further recognition performance.

Standard back propagation (BP) neural network has disadvantages such as slow convergence speed, local minimum and difficulty in definition of network structure. In this paper, an learning vector quantization (LVQ) neural network classifier is established, then it is applied in pattern classification of two-dimensional vectors on a plane. To test its classification ability, the classification results of LVQ neural network and BP neural network are compared with each other. The simulation result shows that compared with classification method based on BP neural network, the one based on LVQ neural network has a shorter learning time. Besides, its requirements for learning samples and the number of competing layers are also lower. Therefore it is an effective classification method which is powerful in classification of two-dimensional vectors on a plane.

This paper examines the utility of audio-visual speech for the two related tasks of speech and speaker recognition. A study of the confusion that exists between speaker and speech elements was performed to show that principal component analysis (PCA) based visual speech is considerably better for the task of speaker recognition than for speech. Decision fusion speech and speaker recognition engines were also tested under various levels of acoustic degradation to find that the optimal fusion configuration for speaker recognition was substantially different than that for speech. These results highlight the problem of employing similar visual features for both speech and speaker recognition.