Showing papers on "Mahalanobis distance published in 1991"

Population Definition, Sample Selection, and Calibration Procedures for Near Infrared Reflectance Spectroscopy

Abstract: Near infrared spectroscopy relies heavily on the collection of an appropriate population of samples for calibration and the best mathematical procedure to obtain the most accurate calibration. The purpose of this study was to evaluate two algorithms (CENTER and SELECT) for defining the population and selecting samples for calibration. The selected samples were used to compare modified partial least squares regression (MPLSR) with modified stepwise regression (MSR) calibration method. The algorithms were developed to (i) establish the boundaries of a population of samples in terms of the standardized Mahalanobis distance (H) from the mean and (ii) select a small, structured set of samples for calibration using the standardized H distance between sample pairs. Two diverse populations of samples were used to test these approaches. Calibrations were performed using MPLSR and MSR. A standardized H distance of 3.0 from the mean was used as a boundary for excluding spectral outliers from a population, and a minimum standardized H distance between samples of 0.6 provided an adequate number of calibration samples for accurate predictions. Both regression methods provided acceptable validation statistics for crude protein, acid detergent fiber, and in vitro dry matter disappearance. The MPLSR calibration method gave an overall 18% improvement in standard error of performance (SEP) compared with the MSR calibration method.

Splitting of calibration data by cluster analysis

Abstract: The topic of the present paper is the splitting of calibration data into subgroups with improved linearity in each group. The method proposed is based on a criterion which is a weighted average of a Mahalanobis distance and a squared regression residual. The algorithm used to find the solution is based on fuzzy clustering. Two examples are given to illustrate the theory.

Parallel and adaptive clustering method suitable for a VLSI system

Abstract: The authors propose a two-functional network in which adaptive networks are implemented for sophisticated recognition and clustering. In the first subnetwork, self-organized clustering is realized. The clustering is based on Mahalanobis distance. The result of the first subnetwork becomes a vector of similarity values between a given input pattern and all patterns of cluster nodes. The second subnetwork consists of nodes associated with specific labels. All connections between the label nodes of the second functional network and the cluster nodes of the first functional network are determined by supervised learning. Every calculation is executed in parallel and pipelined forms. In addition, the proposed network is experimentally shown to provide good performance. In particular, it is shown that handwritten letters can be accurately recognized by using this network. >

Spectral analysis of peach surface defects

Abstract: The spectral reflectance of peach surface defects (bruise, cut, scar, scale, brown rot, wormhole) in the 350 to 1200 nm range was measured. Sorting criteria which were suitable for use with a machine vision system were developed and evaluated based on their potential for discriminating between defective and normal peach surface. Separability was calculated in terms of the Mahalanobis distance between the two classes. Sorting criteria which were based on ratios of wave-lengths had good separability for some defects, but multivariate sorting criteria with at least one unnormalized component were required for good class discrimination for all peach defect types at a fixed set of wavelengths. A three-wave-length criterion with spectral sensitivities centered at 650, 720, and 815 nm had the highest Mahalanobis distances for the types of peach defects tested.

Profile analysis of the Wechsler intelligence scales: a new index of subtest scatter.

Abstract: A statistical method of profile analysis originally suggested by Huba (1985, 1986), the Mahalanobis distance, D2, is described and discussed with reference to its use with the Wechsler scales. Examples are given to illustrate the properties of D2 and two BASIC computer programs are presented that will calculate D2; one for the WAIS-R and the other which can be readily adapted to the test selection of the user's choice. It is argued that D2 is a useful addition to existing methods of statistical analysis of the Wechsler scales.

Application of neural computation to sound analysis for valve diagnosis

Abstract: Describes a method of compressor valve diagnosis in which a neural network was applied to analyze valve plate sound. A neural network approach to pattern recognition was considered. The system reduces the load of neural networks in function representation and function learning by using appropriate preprocessing methods, for instance the Mahalanobis generalized distance. The resulting system requires only a small-scale neural network, is capable of extracting the invariant features, and attains an extremely high recognition accuracy. >

Using Generalized Distances in Classification of Groups

Abstract: The Mahalanobis generalized distance can advantageously be used to achieve the hierarchical clustering of groups of individuals. With a set of nearly 12,000 biometrical data, comprising populations of 14 different species of clover, we tried four methods to cluster those populations, in order to compare their results and to see whether the numerical classification obtained agrees with the botanical taxonomy. One of those methods is a conventional hierarchical clustering technique, based upon the Euclidean distances between the means of the populations, while the three other methods make use, with an increasing degree of complexity, of the generalized distances. These methods gave obviously better results.

A comparison of clustering procedures based on multiple traits in Gerbera and Dahlia

Abstract: The multivariate methods, namely, principal component analysis (PCA), Tocher's technique, and an iterative method of clustering based on successive reallocation of elements using Mahalanobis D2 statistic, were compared on the basis of average intracluster distance and a ratio index of homogeneity of clusters utilizing the intra- and intercluster distances. The results from two different populations under study showed a superior performance of the iterative method. The optimum clusters formed by the iterative method were uniformly more homogeneous and unique as compared to the clusters obtained by the other two procedures. In addition to these advantages, the iterative method can be easily computerised and is useful in handling the large number of genotypes generated in breeding programmes, whereas the other two methods need a lot of manual labour in addition to their computer analysis.

An inductive approach to remote sensing in forestry

Abstract: A rationale for using inductive methodology in remote sensing applied to monitoring of forest resources is given. The approach is tested using a Landsat TM scene of a part of eastern Norway. The image is classified using cluster analysis and the result of the classification is evaluated using Mahalanobis distances. Problems related to the use of statistical verification, Mahalanobis distances in particular, is discussed. A summary of the results with regard to practical applications is given at the end of the article.

An adaptive multivariate signed-rank test for the one-sample location problem

Abstract: A multivariate affine-invariant adaptive test procedure is proposed for the one-sample location problem. The procedure suggested uses the multivariate sign test based on interdirections suggested by Randles, a multivariate signed-rank procedure suggested by Peters and Randles, and a light-tailed version of the signed-rank procedure. A selection statistic constructed from univariate Mahalanobis distances is used to choose the appropriate sign or signed-rank procedure yielding a large sample test which performs well for a broad class of distributions. The performance of the adaptive procedure is assessed via Monte Carlo simulation results.

Speaker matching system

Abstract: PURPOSE:To perform secure speaker matching not for short-period data, but for data with a period difference by using a BPF parameter and an LPC cepstrum which are relatively tolerant to a secular change in combination, and varying their decision conditions according to the period difference. CONSTITUTION:A speaker voices words of the same document, etc., as that in registration and similar DP matching with a standard pattern is performed by a BPF analysis and an LPC(Linear Predictive Coefficient) analysis. Then the similarity between the input pattern after the DP matching and the standard pattern registered in registration mode is found according to a frame, a BPF number, etc. Then the Mahalanobis generalized distance (D1, D2) is compated form the distribution for the person himself and the speaker distribution of the input data; when D1 > D2, the pattern is decided as that of a false speaker and rejected, but when D1 < D2, the pattern is judged as that of the person himself and accepted.

Comparative study of texture measurements for cellular organelle recognition

Abstract: The ability of four methods to perform automatic texture discrimination of three cellular organelles (nucleus, mitochondria and lipid droplets) from autoradiographic images is investigated. The four methods studied are the first-order statistics of the gray-level histogram, the gray-level difference method, the gray-level run length method, and the spatial gray-level dependence method. The influence of parameters like the number of features, the number of gray-level classes, the orientation and step size of the analysis, and the effect of preprocessing the images by histogram equalization and image reduction were also analyzed to optimize the performance of the methods. The nearest neighbor pattern recognition algorithm using the Mahalanobis distance was used to evaluate the performance of the methods. First, a training set of 30 samples per organelle was chosen to train the classifier and to select the best discriminant features. The probability of error was estimated with the leave-one-out method and the results are expressed in percentage of correct classifications. The study shows that features extracted using the spatial gray-level dependence method were the most discriminate ones. The best features set was then applied to a test population of 734 cellular organelles to differentiate the three classes. Correct classifications occurred for 95% of cases, which indicates that it is possible to achieve a semi-automatic analysis of autoradiographic images.