Showing papers on "Mahalanobis distance published in 1997"

A Measure of DNA Sequence Dissimilarity Based on Mahalanobis Distance Between Frequencies of Words

Abstract: SUMMARY A number of algorithms exist for searching genetic databases for biologically significant similarities in DNA sequences. Past research has shown that word-based search tools are computationally efficient and can find similarities or dissimilarities invisible to other algorithms like FASTA. We characterize a family of word-based dissimilarity measures that define distance between two sequences by simultaneously comparing the frequencies of all subsequences of n adjacent letters (i.e., n-words) in the two sequences. Applications to real data demonstrate that currently used word-based methods that rely on Euclidean distance can be significantly improved by using Mahalanobis distance, which accounts for both variances and covariances between frequencies of n-words. Furthermore, in those cases where Mahalanobis distance may be too difficult to compute, using standardized Euclidean distance, which only corrects for the variances of frequencies of n-words, still gives better performance than the Euclidean distance. Also, a simple way of combining distances obtained at different n-words is considered. The goal is to obtain a single measure of dissimilarity between two DNA sequences. The performance ranking of the preceding three distances still holds for their combined counterparts. All results obtained in this paper are applicable to amino acid sequences with minor modifications.

Mass detection in digital radiologic images using a two stage classifier

Abstract: A computer-implemented method of identifying suspect masses in digital radiologic images and a system for computer-aided diagnosis of such images in which the images are thresholded at a large number of threshold levels to discriminate spots and a two stage classifier is applied to the spots. The first classification stage applies multiples rules, predetermined from a training set of images, to a relatively computationally inexpensive set of initial features, namely area, compactness, eccentricity, contrast, and intensity variance for each spot. More computationally expensive features, namely edge orientation distribution and texture features, are computed only for spots that are accepted by the first classification stage to points for these spots in an expanded feature space. In the second classification stage, these points are classified as true positives or false positives in dependence on which means of a plurality of clusters of true positives and a plurality of clusters of false positives, predetermined from the training set, are nearest in Mahalanobis distance.

Rotated texture classification by improved iterative morphological decomposition

Abstract: An improved algorithm based on iterative morphological decomposition (IMD) proposed by Wang et al. (1993) is described. The proposed algorithm requires less computation than the original IMD algorithm. The improved iterative morphological decomposition (IIMD) is compared with granulometric moments, multiresolution rotation-invariant SAR (MRRISAR) models and multichannel Gabor filters. It is found that IIMD is superior to granulometric moments and MRRISAR in rotated texture classification. IIMD may also perform better than multichannel Gabor filters by employing many different kinds of structuring elements. In the study, three kinds of pseudo rotation-invariant structuring elements, namely the disc, octagon and square, as well as a line structuring element are tested. Since the line structuring element is rotation-variant in nature, the image is rotated to different orientations of equal angular separation to find a set of primitive features. A Fourier transform is then applied to convert these features to rotation-invariant. An accuracy rate as high as 96% is achieved in classifying 30 classes of textured images in the experiment. It is also demonstrated that using both the normalised variance and the mean can give better classification accuracy rate than using both the variance and the mean when classified by simplified Bayes or Mahalanobis distance measure.

An adaptive, on line, statistical method and apparatus for motor bearing fault detection by passive motor current monitoring

Abstract: A motor current signal is monitored during a learning stage and divided into a plurality of statistically homogeneous segments representative of good operating modes. A representative parameter and a respective boundary of each segment is estimated. The current signal is monitored during a test stage to obtain test data, and the test data is compared with the representative parameter and the respective boundary of each respective segment to detect the presence of a fault in a motor. Frequencies at which bearing faults are likely to occur in a motor can be estimated, and a weighting function can highlight such frequencies during estimation of the parameter. The current signal can be divided into the segments by dividing the current signal into portions each having a specified length of time; calculating a spectrum strip for each portion; and statistically comparing current spectra of adjacent ones of the strips to determine edge positions for the segments. Estimating the parameter and the boundary of each segment can include calculating a segment mean (the representative parameter) and variance for each frequency component in each respective segment; calculating a modified Mahalanobis distance for each strip of each respective segment; and calculating the modified Mahalanobis mean and the variance for each respective segment. Each modified Mahalanobis mean can form a respective radius about a respective segment mean to define a respective boundary.

A handwritten character recognition system using modified Mahalanobis distance

Abstract: An evaluation distance function is one of the most important factors that influences the accuracy of a handwritten character recognition system. Various evaluation distance functions have been proposed and investigated theoretically. City block distance, Euclidean distance, weighted Euclidean distance, sub-space method, multiple similarity method, Bayes decision method and Mahalanobis distance are known typical distance functions. Although observing the performance of each evaluation function in a large-scale handwritten character recognition system is quite important, there has been little research reported on this topic. In this paper, because the emphasis is on how to improve the accuracy of a recognition system, comparison experiments are carried out. The experimental results show that the Mahalanobis distance is the most effective of the seven typical evaluation distance functions. Considering the foregoing result and the properties of distribution on each axis, a modified Mahalanobis distance is proposed to construct a more accurate and faster system. Using the proposed modified Mahalanobis distance as the evaluation function, a recognition rate of 98.24 percent has been achieved for ETL9B, the largest public database of handwritten characters in Japan. In this paper, the behavior of seven typical evaluation functions is studied and a new evaluation distance, called the modified Mahalanobis distance, is proposed based on these results. © 1997 Scripta Technica, Inc. Syst Comp Jpn, 28 (1): 46–55, 1997

Model diagnostics for marginal regression analysis of correlated binary data

Abstract: We propose several diagnostic methods for checking the adequacy of marginal regression models for analyzing correlated binary data We use a parametric marginal model based on latent variables and derive the projection (hat) matrix, Cook's distance, various residuals and Mahalanobis distance between the observed binary responses and the estimated probabilities for a cluster Emphasized are several graphical methods including the simulated Q-Q plot, the half-normal probability plot with a simulated envelope, and the partial residual plot The methods are illustrated with a real life example

Self-organizing geometric certainty maps: a compact and multifunctional approach to map building, place recognition and motion planning

Abstract: In this paper we show how a self-organizing Kohonen neural network can use hyperellipsoid clustering (HEC) to build maps from actual sonar data. Since the HEC algorithm uses the Mahalanobis distance, the elongated shapes (typical of sonar data) can be learned. The Mahalanobis distance metric also gives a stochastic measurement of a data point's association with a node. Hence, the HEC Kohonen can be used to build topographical maps and to recognize its own topographical cites for self-localization. The number of nodes can also be regulated in a self-organizing manner by using the Kolmogorov-Smirnov (KS) test for cluster compactness. The KS test determines whether a node should be divided (mitosis) or pruned completely. By incorporating principal component analysis, the HEC Kohonen can handle problems with several dimensions (3D, time-series, etc.). The HEC Kohonen is also multifunctional in that it accommodates compact geometric motion planning and self-referencing algorithms. It can also be used to solve a host of other pattern recognition problems.

Comments on "A self-organizing network for hyperellipsoidal clustering (HEC)" [with reply]

Abstract: In the above paper by Mao-Jain (ibid., vol.7 (1996)), the Mahalanobis distance is used instead of Euclidean distance as the distance measure in order to acquire the hyperellipsoidal clustering. We prove that the clustering cost function is a constant under this condition, so hyperellipsoidal clustering cannot be realized. We also explains why the clustering algorithm developed in the above paper can get some good hyperellipsoidal clustering results. In reply, Mao-Jain state that the Wang-Xia failed to point out that their HEC clustering algorithm used a regularized Mahalanobis distance instead of the standard Mahalanobis distance. It is the regularized Mahalanobis distance which plays an important role in realizing hyperellipsoidal clusters. In conclusion, the comments made by Wang-Xia together with this response provide some new insights into the behavior of their HEC clustering algorithm. It further confirms that the HEC algorithm is a useful tool for understanding the structure of multidimensional data.

Optimal trade-off distance classifier correlation filters (OTDCCFs) for synthetic aperture radar automatic target recognition (SAR ATR)

Abstract: Recent developments in optimal trade-off based composite correlation filter methods have improved the recognition and classification of an object over a range of image distortions. We extend the capability of the distance classifier correlation filter introduced by Mahalanobis et al by using he optimal trade-off between different correlation criteria. These correlation filters can be used for the automatic target cueing or recognition of synthetic aperture radar (SAR) images. In this paper we will present results of designing these distortion-tolerant filters with simulated SAR imagery and testing with simulated SAR target images inserted into real SAR backgrounds.

Color Gamut Mapping based on Mahalanobis Distance for Color Reproduction of Electronic Endoscope Image under Different Illuminant.

Abstract: In this paper, a color difference is defined instead of the conventional color differences to evaluate gamut-mapping techniques for electronic endoscope images reproduced on CRT under environmental illuminant. This color difference is defined from Mahalanobis distance by using covariance matrices for differences of metric lightness, chroma and hue angle. The covariance matrices for endoscope images were obtained by psychophysical experiments. We compared the resultant matrices with those for natural scenes, and found the color difference was customized for the electronic endoscope images well.

Using a hyper-ellipsoid clustering Kohonen for autonomous mobile robot map building, place recognition and motion planning

Abstract: We show how a self-organizing Kohonen neural network using hyperellipsoid clustering (HEC) can build maps from actual sonar data. With the HEC algorithm we can use the Mahalanobis distance to learn elongated shapes (typical of sonar data) and obtain a stochastic measurement of data-node association. Hence, the HEC Kohonen can be used to build topographical maps and to recognize its own topographical cues for self-localization. The number of nodes can also be regulated in a self-organizing manner by measuring how well a node models the statistical properties of its associated data. This measurement determines whether a node should be divided (mitosis) or pruned completely. Because fewer nodes are needed for an HEC Kohonen than for a Kohonen that uses only Euclidean distance, the data size is smaller for the HEC Kohonen. Relative to grid-based approaches, the savings in data size is even more profound. By incorporating principal component analysis (PCA), the HEC Kohonen can handle problems with several dimensions (3D, time-series, etc.). The HEC Kohonen is also multifunctional in that it accommodates compact geometric motion planning and self-referencing algorithms. It can also be generalized to solve other pattern recognition problems.

Multiple-time-point dissolution specifications for a sampling acceptance plan

Abstract: In dissolution testing, multiple dissolution measurements at specific time points are used to obtain the dissolution characteristics for most extended-release and some immediate-release drug products. This paper presents a general procedure for defining specifications by a multivariate confidence region or by simultaneous confidence limits on the dissolution values of individual time points. The confidence regions and simultaneous confidence limits were estimated using two approaches: the first approach assumed a multivariate normal distribution on the multiple dissolution values and the second approach used the bootstrap resampling method. The multivariate confidence region was constructed using the Hotelling's T2 statistic (equivalently, Mahalanobis distance D2), and the simultaneous confidence limits were based on the Mahalanobis statistic as well as on the Bonterroni adjustment. The Mahalanobis simultaneous confidence limits specification criteria have the overall false out-of-specification rate too low in both parametric and bootstrap approaches.

Validation of an algorithm for atrial fibrillation and flutter diagnosis

Abstract: In this work, the authors developed a methodology to evaluate an algorithm (AAFF) for atrial fibrillation (AF) and flutter (AFL) detection and discrimination. Atrial electrical activity in pathologies such as AF and AFL are difficult to characterize quantitatively. The diagnoses provided by AAFF were compared to those established by 8 clinicians at different expertise levels in cardiology, and the MIT-BIH database annotations. The concordance between diagnoses supplied by a pair of experts was studied with the measure of distances between their diagnoses. The methods used to compute distances were: Euclidean distance, Mahalanobis distance and City-block distance. Using the resulting matrices of distances between experts, cluster analyses were carried out to classify AAFF among human experts. The results show that differences between diagnoses made by AAFF and those made by some specialists were smaller than differences between some specialists themselves.

Extraction of features for classification of impulse radar measurements

Abstract: The present paper addresses the problem of extracting features for classification purposes. A vector valued sample is to be classified to one of a number of classes with known distributions using the Bayes decision rule. The complexity of the classifier depends on the dimension of the vectors; thus it is of interest to keep this dimension as small as possible. One way to reduce the dimension is to apply a linear transformation on data. This transformation should be chosen so that no 'essential' information is lost. There are several suggestions on how this concept should be defined. We study a measure of class separability defined as the mean of all interclass Mahalanobis distances. The method to be presented, however, applies to all weighted quadratic distance measures. The validity of the proposed transformation is justified by applying the transformation to both Monte Carlo simulated data and to actual measured data. The measured data come from an impulse radar system with the purpose of classifying buried objects. The proposed transformation is shown to outperform the well known principal component analysis (PCA).© (1997) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

The Font Recognition of Printed Hangul Documents

Abstract: The main focus of this paper is the recognition of printed Hangul documents in terms of typeface, character size and character slope for IICS(Intelligent Image Communication System). The fixed-size blocks extracted from documents are analyzed in frequency domain for the typeface classification. The vertical pixel counts and projection profile of bounding box are used for the character size classification and the character slope classification, respectively. The MLP with variable hidden nodes and error back-propagation algorithm is used as typeface classifier, and Mahalanobis distance is used to classify the character size and slope. The experimental results demonstrated the usefulness of proposed system with the mean rate of 95.19% in typeface classification. 97.34% in character size classification, and 89.09% in character slope classification.

ML parameter estimation and minimum distance classifier for texture analysis using wavelet transform

Abstract: This work concentrates on comparing the performance of the minimum distance classifier and maximum-likelihood classifier for texture analysis. A tree-structured wavelet transform has been used for extracting the features and the comparison is based on the correct classification percentage. The results indicate that the maximum-likelihood classifier performs marginally better than the mahalanobis distance for some feature sets. The Euclidean distance did not prove to be powerful in distinguishing the textures. The performance of various orthogonal wavelet transforms have also been compared in order to find out the best wavelet for each of the classifiers considered.

Comments on "A Self-organizing Network for Hyperellipsoidal Clustering (HEC)"

Abstract: In the above paper, 1 Mahalanobis distance is used instead of Euclidean distance as the distance measure in order to acquire the hyperellipsoidal clustering. We prove that the clustering cost function is a constant under this condition, so hyperellipsoidal clustering cannot be realized. This letter also explains why the clustering algorithm developed in the above paper can get some good hyperellipsoidal clustering results.

Articles identifying system

Abstract: PROBLEM TO BE SOLVED: To identify tablets even if the front/rear of the tablets cannot be specified by assuming an article belongs to an inputted category when at least one of two calculated Mahalanobis' distance is within a prescribed threshold distance. SOLUTION: A distance calculation expression derivation processing 105 obtains 'Mahalanobis' distance' calculation expressions for every group (A or B) for every tablet brand by using the feature vector of many tablet samples which are inputted up to the present. In an identification phase, the unknown table and a brand group, whose corresponding is not known, are set to be inputs. In an identification processing 107, identification is executed by using the general-distance of corresponding to the reference vectors of the groups A and B, which is calculates in the general distance calculation processing 106. When the Mahalanobis' distance corresponding to either the group A or the group B is within a prescribed threshold distance, the tablet is judged to correspond to the inputted brand. When it is not, the tablet is judged so that it does not correspond to the brand and the result is outputted.

Variations on principal components and subspace projection for remote hyperspectral classification

Abstract: This paper discusses recently developed algorithms for the classification of pixels in hyperspectral images, used in conjunction with a library of hyperspectral hemispherical reflectance data measured in the laboratory and partitioned into usable classes of materials. The algorithms are based upon functions of the principal components of the class covariances and the corresponding null spaces, and the underlying measures used in the classification statistics are similar to Mahalanobis distances. The algorithms can be used as stand-alone processing or combined with spatial and temporal algorithms n a higher level system of hyperspectral image processing. The nature of the classification algorithms and the database will be discussed, with particular attention being paid to issues specific to this approach. The basic performance of the classifier algorithms will be demonstrated using modified laboratory data. The applicability of orthogonal subspace projection methods to problems inherent in remote sensing using hyperspectral invisible and IR data will be emphasized, while specifically dealing with the compensation for inaccuracies in necessary estimates of atmospheric attenuation and target temperature. Preliminary results of classification of field collected hyperspectral data will also be presented, and ongoing and future work in hyperspectral classification described.

Wavelet transform algorithm for automated radar-to-IR track-file association

Abstract: The association of target track files established and maintained by physically separated sensors based solely upon metrics data does not provide sufficient performance to meet some weapon system requirements. Association algorithms that match tracks based on similarities in the Fourier power spectra derived from the targets' signatures are routinely employed to improve post-mission confidence in track associations made between airborne IR tracking sensors and ground or ship-based radar frequency (RF) sensors that view common target suites during live exercises. The problem with Fourier techniques is that long viewing times are required to obtain usable power spectral density estimates; our mission scenarios impose a time constraint that is about one order of magnitude less. Faster algorithms are required for real-time embedded interceptor and BMC4I applications. This paper documents preliminary results we have obtained using a prototype, wavelet-based algorithm to automate the rapid, one-to-one association of RF and IR target tracks using features extracted from short signature histories maintained in the track files. The primary physical phenomenon that is exploited by the algorithm is the frequency content in the target signature, which is induced by the target's body dynamics. No a priori knowledge of the expected signature dynamics is assumed. Given a group of targets that are observed by two or more sensors, each sensor independently establishes and maintains its own track files while viewing the targets from physically isolated platforms. The sensors may have different sample rates and may operate in different regions of the RF spectrum; concurrent viewing is not required. The algorithm automatically detects and extracts signature glints as a preprocessing step, saving the glint as a preprocessing step, saving the glint data in separate 'channels' for further processing and employs wavelet shrinkage to attenuate any white noise that may be present in the signature data. Individual sensor track files are treated as separate pattern classes within an adaptive, statistical pattern recognition framework. The class feature vectors are formed from the magnitudes of the wavelet packet crystal coefficients. Clustering transformations are applied to each 'class', Fisher's linear discriminant is employed to minimize the intraclass scatter while maximizing the interclass scatter, and a modified Mahalanobis distance is then used as a metric to quantify the similarity between each possible pair of classes. Evidence is provided that suggests the wavelet packet crystal energies used by the association algorithm may be of some utility in the detection of closely spaced objects. The algorithm and the analyses are discussed within the context of a two-sensor scenario, but the algorithm is equally applicable to multiple sensor applications.© (1997) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

A wavelet transform algorithm for Radar-to-IR track file association

Abstract: The association of target track files established and maintained by physically separated sensors based solely upon metrics data (i.e. predicted target positions, velocities etc.) does not provide sufficient performance to meet some weapon system requirements. Association algorithms that match tracks based on similarities in the Fourier power spectra derived from the targets' signatures are routinely employed to improve post-mission confidence in track associations made between airborne infrared (IR) tracking sensors and ground or ship-based radar frequency (RF) sensors that view common target suites during live exercises. The problem with Fourier techniques is that long viewing times are required to obtain usable power spectral density (PSD) estimates (10 to 20 s is typical); our mission scenarios impose a time constraint that is about one order of magnitude less. Faster algorithms are required for real-time embedded interceptor and BMC 4 I applications. This paper documents preliminary results we have obtained using a prototype, wavelet-based algorithm to automate the rapid, one-to-one association of RF and IR target tracks using features extracted from short signature histories maintained in the track files. The primary physical phenomenon that is exploited by the algorithm is the frequency content in the target signature, which is induced by the target's body dynamics. No a priori knowledge of the expected signature dynamics is assumed. Given a group of targets that are observed by two or more sensors, each sensor independently establishes and maintains its own track files while viewing the targets from physically isolated platforms. The sensors may have different sample rates and may operate in different regions of the RF spectrum ; concurrent viewing is not required. The algorithm automatically detects and extracts signature glints as a preprocessing step, saving the glint data in separate channels for further processing and employs wavelet shrinkage to attenuate any white noise that may be present in the signature data. Individual sensor track files are treated as separate pattern classes within an adaptive, statistical pattern recognition framework. The class feature vectors are formed from the magnitudes of the wavelet packet crystal coefficients. Clustering transformations are applied to each class:' Fisher's linear discriminant is employed to minimize the intraclass scatter while maximizing the interclass scatter, and a modified Mahalanobis distance is then used as a metric to quantify the similarity between each possible pair of classes (target tracks). Evidence is provided that suggests the wavelet packet crystal energies used by the association algorithm may be of some utility in the detection of closely spaced objects (CSOs).The algorithm and the analyses are discussed within the context of a two-sensor scenario. but the algorithm is equally applicable to multiple sensor applications.