Learning Vector Quantization (LVQ)

system is proposed. The feature extraction techniques employed in system extract Mel frequency cepstral coefficient (MFCC), delta mel frequency cepstral coefficient (DMFCC) and format frequency. The feature selection is done using hybrid model of particle swarm optimizatiom (PSO) and Genetic Algorithm (GA). We have used Learning Vector Quantization (LVQ) artificial Neural Network classifiers. The speech database consists of 40 speakers (20 males+ 20 females) speech utterance. The speech utterance is recorded for a specific sentence in three different languages viz. “Now this time you go” (in English), “Adhuna Asmin Twam Gachh “(in Sanskrit), “Ab Iss Baar Tum Jao” (in Hindi). Total word for this purpose is 14 including 4 for Sanskrit and 5 Hindi and English. The average identification rate 79.99% is achieved when the Network is trained by LVQ and it shows 80.52% when LVQ is trained using hybrid PSO-GA model.

Recently some work has gone into the detection and classification of brain tumors. One such effort in the domain of meningiomas was carried out by Lessman et al. [1] who studied the problem of content-based visualization of meningioma images to aid in characterization of the database contents. In this paper, we combine the feature extraction capability of the Discrete Wavelet Packet Transform (DWPT) based upon the selection of the subbands that are most useful for texture classification with classification ability of neural networks. The ability of wavelet packet transform in devising a feature selection algorithm for texture classification was demonstrated in [2, 3]. Our work is motivated by the basis selection paradigm presented in [4] and texture classification using discriminant wavelet packet subbands [5]. Statistical parameters [6] of the selected subbands are used as features for training and detection purposes. Learning Vector Quantization (LVQ) [7] is used as a classifier to discriminate between different kinds of meningiomas. LVQ has been used in the literature to discriminate between images based upon texture [8] and to classify pancreatic tissues based on their texture [9]. The following section describes the methodology of the proposed algorithm in detail. The results are presented at the end with a discussion and future directions.

In machine learning, speech and pattern recognition feature extraction plays a key role. In speech recognition it is the key step so that it is necessary to keep more attention to feature extraction procedure. Feature Extraction reduces the amount of resources to show a large set of the input data and gives the compact view of the input speech signal. It also converts the audio signal into numerical forms, these features values are applied to the Learning Vector Quantization (LVQ) neural network for further procedure. In this proposed project MFCC [11] and LPC [1] are used for the feature extraction technique.

Human is gifted by god with five senses – sight, hearing, touch, smell and taste – which humans use to perceive their environment. Out of these five senses, sight is the most powerful. Image Contrast Enhancement with brightness preserving is a simple, effective and most widely used area among all digital image processing techniques. The goal of brightness preserving and contrast enhancement in general is to provide a more appealing image and clarity of details. These enhancements are intimately related to different attributes of visual sensation. In this paper we propose a method of image enhancement using Learning Vector Quantization for feature enhancement. Result shows a significant performance improvement by applying LVQ. Proposed method results generate better values of Absolute Mean Brightness Error (AMBE) and Peak Signal to Noise Ratio (PSNR) than other Histogram Equalization (HE) method.

can be used as classification’s input [4]. Agustina use the ratio of the nucleus to the cytoplasm and produces an accuracy of 78% for abnormal cells image [5]. Meutia use the image intensity values from segmented cervical cell image to classify cervical cells using Learning Vector Quantization (LVQ) method which produces an accuracy of 82% [6]. LVQ is one of the ANN models which has a fairly simple architecture but has good performance as a classification engine.

The player’s motivation a significant role in the success of the learning process and of the game for educational purpose. However, not an easy to determine the level of player’s motivation while playing the serious game. To assess the motivation level of player interest, this paper proposes a Motivation Behavior Game (MBG). MBG improves this motivation concept to monitor how players interact with the game. This game employs Learning Vector Quantization (LVQ) for optimizing the motivation behavior input classification of the player. MBG is using teacher’s data to obtain the neuron vector of motivation behavior pattern supervise. Three clusters multi objective target will be classified as; active choice, persistence, and mental effort motivation behavior. In the game play experiments employ 33 respondent players demonstrates that 12.12% of players have high and 6.06% have semi mental effort, 3.03% have high and 3.03% semi persistence, and 66.67% have high and 9.09% low active choice motivation behavior. MBG may provide information to game engine when a player needs help or when wanting a formidable challenge. The game engine will provide the appropriate tasks according to players’ ability. MBG will help balance the emotions of players, so players do not get bored and frustrated. The high interest players will finish the game if their emotions are stable. The players’ interests strongly support the procedural learning in a serious game.

The dimensionality of the data gets increased as the databases in general are modified to suit to the changing requirements of users. As a result, databases will be suffered with increased redundancy, which in turn effects the accuracy and prediction quality of classification tasks. This paper is aimed at increasing the classification performance by eliminating highly correlated and redundant features/attributes through newly identified feature selection methods. An attempt has been made to prove the efficacy of dimensionality reduction by applying LVQ(Learning Vector Quantization) method on benchmark dataset of ‘Lung cancer Patients’. In this paper, the newly identified feature selection methods Correlation and Coefficient of dispersion were adopted for facilitating in dimensionality reduction, removing inappropriate data, improving comprehensibility and learning accuracy.

This paper has presented a new over-sampling method using codebooks obtained by Learning Vector Quantization. In general, even when an existing SMOTE is applied to a biomedical dataset, it is still difficult to estimate proper borderlines between classes. In order to tackle this problem, we have proposed to generate synthetic samples using code- books obtained by the learning vector quantization. The experimental results on eight real-world benchmark datasets have shown that the proposed over-sampling method gen- erates useful synthetic samples for the classification of imbalanced biomedical data. It is expected that the proposed over-sampling method is basically compatible with basic classification algorithms and the existing over-sampling methods. In addition, experi- ments on datasets for β -turn types prediction show our proposed method has improved prediction of β -turns type IV and VIII.

fonts, syntax, weather and environment, spacing etc. Therefore, most of the previous prescripts could not apply for all the countries in the real world, all the environments, all types of the Kingdom of Saudi Arabia Vehicle License Plate. The Kingdom of Saudi Arabia Vehicle license plate localization and recognition system today will be required to operate robustly in environments with intricate backgrounds and light intensity variations. To deal with such problems, researchers have proposed in this paper uses Artificial Neural Network techniques [6] for character recognition, using Learning Vector Quantization Neural Network [7]. Their results are compared based upon their completeness in the character recognition. The efficiency of the System can be further improved by increasing the number of fonts for training Learning Vector Quantization Artificial Neural Networks.

Based on this, our goal in this document is to investigate effective methods that produce accurate quantizations with data samples. One of the most popular procedure is Lloyd’s algorithm (see Lloyd, 2003) sometimes refereed to as batch k-means. A convergence theorem for this algorithm is provided by Sabin and Gray (1986). Another celebrated quantization algorithm is the competitive learning vector quantization (CLVQ), also called on-line k-means. The latter acronym outlines the fact that data arrive over time while the execution of the algorithm and their characteristics are unknown until their arrival times. The main difference between the CLVQ and the Lloyd’s algorithm is that the latter run in batch training mode. This means that the whole training set is presented before performing an update, whereas the CLVQ algorithm uses each item of the training sequence at each update.

In computer aided diagnosis, pre-processing, segmentation, feature extraction and classification are the steps involved. For segmentation the boundary must be defined to get regional information inside the boundary. Ground Glass Opacity (GGO) has ill-defined boundary. Hence, there is a necessity to optimize the boundary of GGO. Once the boundary is optimized, feature extraction and classification of malignant and benign of the particular GGO becomes easy. Using Distance Regularized Level Set Evolution (DRLSE) and Active contour without edges independently, the contour is grown and compared with the model image. The model image was created using the expertise of a Radiologist, applying Mattes Mutual information method. The contour which gives maximum mutual information is concluded as the optimized boundary. Wavelet transform has already been proven its application in identifying pits and cracks of corrosion metals. The same analogy is applied in GGO so that the partly solid and liquid GGOs can be precisely classified as malignant or benign. After optimizing the ill-defined boundary of GGO features of wavelet transformation were extracted along with textural features of mean and variance. Skewness and kurtosis were neglected since they were negligibly small. It is shown that on comparing the growing contour with model image using mattes mutual information, the DRLSE method shows greater results, without leaking compared to Active contour without edges. After extraction of features from wavelet transformation and textural features classification as malignant and benign was done using learning vector quantization (LVQ). On finding the optimized boundary, it is easier to classify the ground glass opacity as diffuse finding or local finding. Hence taking two images where one is taken as malignant and other as benign classification was done as benign and benign by the classifier. The malignant have been identified as benign due to minimum number of images used in training.

ABSTRACT: It is crucial to quantify location of sensor in the dam, for this purport numbers of ,the sensor are placed at sundry locations in the dam. Central water and Power Research Station (CWPRS) provided data of Dam sensor which utilized in this project to train the Neural Network. Identification of sensor location is the arduous task if the quandary is arrived in the control room, because of Mismatching in the wire, Noise. For solving this quandary Artificial Neural Network is utilized. Compared with some traditional methods of Artificial Neural Network has lots of advantages in dealing with the highly non-linear quandary. Neural Network has Back propagation (BP), Conjugate Gradient Method, Learning Vector Quantization (LVQ) Algorithms and Levenberg-Marquardt (LM), Neuron by Neuron (NBN), Architectures. Artificial Neural Network used to identify exact sensor predicated on erudition about the dataset.

Identically and independently distributed LEEWMA Last value of exponentially weighted moving average LVQ Learning vector quantization MAT Mean x autocomelation 1MGUSUM Multivariate cumul[r]

A Learning Vector Quantization Network (LVQ) has a first competitive layer and a second linear layer. The competitive layer learns to classify input vectors in much the same way as the competitive layers of Self-Organizing Nets. The linear layer transforms the competitive layer's classes into target classifications defined by the user. The classes learned by the competitive layer are referred to as subclasses and the classes of the linear layer as target classes. Both the competitive and linear layers have one neuron per class. LVQ learning in the competitive layer is based on a set of input/target pairs. The second layer needs no learning as the output classes are known for each input pattern [6]. C. Redial Basis Function

Two-stage training [17][22][36][264] is often used for constructing RBF neural networks. At the first stage, the hidden layer is constructed by selecting the center and the width for each hidden neuron using various clustering algorithms. At the second stage, the weights between hidden neurons and output neurons are determined, for example by using the lin- ear least square (LLS) method [22]. For example, in [177][280], Kohonen’s learning vector quantization (LVQ) was used to determine the centers of hidden units. In [219][281], the k-means clustering algorithm with the se- lected data points as seeds was used to incrementally generate centers for RBF neural networks. Kubat [183] used C.4.5 to determine the centers of RBF neural networks. The width of a kernel function can be chosen as the standard deviation of the samples in a cluster. Murata et al. [221] started with a sufficient number of hidden units and then merged them to reduce the size of an RBF neural network. Chen et al. [48][49] proposed a constructive method in which new RBF kernel functions were added gradually using an orthogonal least square learning algorithm (OLS). The weight matrix is solved subsequently [48][49].

139 | P a g e important, yet usually hidden, information from the raw data. Even more important is the mixing of suitable feature extractor and pattern classifier such that they can operate in coordination to make an effective and efficient CAD system. Statistical features [2], time-domain features [3, 4] and cross-correlation based frequency–domain features [5, 6] were used by different researchers for features extraction. In [7, 8], wavelet transform were used for features extraction, and the authors of [9] used Lyapunov exponents for the same purpose. For classification purpose different classifier [2 ,3, 5, 6] were used by the researchers. 21 features vector are extracted from cross-wavelet and cross-wavelet coherence spectrum. In this paper ,LVQ classifier is used for classify normal and abnormal beats. A brief of the topics on which the different sections are concerned is given as follows. In section 2 information about ECG datasets used in our work is given. Section 3 gives an idea of cross-wavelet transform of two time domain signals. Section 4 deals with the idea of Learning Vector Quantization and how it works. Performance of the classifier scheme is shown in section 5. Section 6 describes the results obtained in this work and future research scope using EEG signal.

feature extraction is done to determine the type of brain tumor, segmentation to locate tumor. In method used by Mohd Fauzi Bin Othman, et al[6], performed classification of brain tumor using wavelet based feature extraction method and Support Vector Machine (SVM), Accuracy of only 65% was obtained. Salim Lahmiri et al [4], proposed approach shows that feature extraction from the LH (Low-High) and HL (High-Low) sub-bands using first order statistics has higher performance than features from LL (Low-low) bands. Ahmed kharrat et al [5], in this paper A Hybrid Approach for Automatic Classification of Brain MRI Using Genetic Algorithm and SVM is used. This paper proposes a genetic algorithm and SVM based classification of brain tumor. It is found that, Gabor filters are poor in performance due to their lack of orthogonality that results in redundant features at different scales or channels. Bhagwat et al[1], in this paper comparison of K-means, Fuzzy C-means and Hierarchical clustering algorithms for detection of brain tumor is carried out, it observed that K-means algorithm produce more accurate result compared to Fuzzy c-means and hierarchical clustering. Pankaj Sapra, et al [11]. This proposes a modified Probabilistic Neural Network (PNN) model that is based on learning vector quantization (LVQ) with image and data analysis and manipulation techniques is proposed to carry out an automated brain tumor classification using MRI-scans. Simulation output showed that PNN system presented and successfully handle the process of brain tumor classification in MRI image with 100% accuracy. From the literature survey, we can say that, various researchers worked on classifying MR brain images into normal and abnormal. Discrete Wavelet Transform is found to be an important tool in decomposing the images into different levels of resolution, from which the significant features can be extracted. Conventional neural networks are inferior to PNN in terms of its accuracy in classifying brain tumors. Hence a wavelet and co occurrence matrix method based texture feature extraction and Probabilistic Neural Network for classification and segmentation by k means has been used in this method.

In this paper the neural network algorithm is proposed for diseased plant leaf classification. The neural network techniques such as feed forward neural network (FFNN), learning vector quantization (LVQ) and radial basis function network (RBF) were tested for two different diseased leaf image classifications such as bean and bitter gourd leaves. The performance is measured using classification parameters such as Accuracy, Precision, Recall ratio and F_measure. With these four parameters the performance is analyzed and based on the analysis the FFNN classification approach provides better result.

Abstract—It is known that learning methods of fuzzy mod- eling using vector quantization (VQ) and steepest descend method (SDM) are superior in the number of rules to other methods using only SDM. There are many studies on how to realize high accuracy with a few rules. Many methods of learning all parameters using SDM are proposed after determining initial assignments of the antecedent part of fuzzy rules by VQ using only input information, and both input and output information of learning data. Further, in addition to these initial assignments, the method using initial assignment of weight parameters of the consequent part of fuzzy rules is also proposed. Most of them are learning methods with simplified fuzzy inference, and little has been discussed with TS (Takagi Sugeno) fuzzy inference model. On the other hand, VQ method with supervised learning that divides the input space into Voronoi diagram by VQ and approximates each partial region with a linear function is known. It is desired to apply the method to TS fuzzy modeling using VQ. In this paper, we propose new learning methods of TS fuzzy inference model using VQ. Especially, learning methods using VQ with the supervised learning are proposed. Numerical simulations for function approximation, classification and prediction problems are performed to show the performance of proposed methods.

In this SVM method, w and b are the key for generating SVM model. For optimizing hyperplane with w and b, gradient descent is used as an optimization algorithm. It is a simple and ancient method for solving optimization problems. For finding a minimum of function, w and b need to be optimized constantly with appropriate direction and step size. The direction is the gradient of the function, and the step size is the learning-rate. Appropriate learning-rate is essential toward the minimum of function, and it gradually diminishes for converging when function value approximates a minimum. By using this method, the optimal hyperplane can be obtained conveniently.