Multiple Classification Technique Research Articles

Performance Comparison of SAGE and MUSIC for Channel Estimation in Direction-Scan Measurements

In this paper, the performances of a space-alternating generalized expectation-maximization (SAGE) algorithm and a MUltiple Signal Identification Classification (MUSIC) technique are compared when they are used to estimate the parameters of multipath components in a propagation channel from measurement data acquired by stepping a horn antenna in directions. The impacts of the horn antenna’s half-power beamwidth (HPBW) and the stepping strategy on root mean squared estimation error and the intrinsic resolution of the estimators are investigated through simulations. The results demonstrate that the performances of both methods rely on the settings of the direction-scan strategy and the antenna HPBW. In general, the MUSIC algorithm, which owns a superior resolution, is less sensitive to the scan configuration and more suitable for scenarios of high signal-to-noise power ratio compared with the SAGE algorithm. Measurement data collected at 13–17 GHz is processed by using both algorithms. The results show that the MUSIC technique can be used to provide similar results with those of the SAGE algorithm at a time consumption significantly reduced.

FINGERPRINT RECOGNITION USING MULTIPLE CLASSIFIER SYSTEM

In this paper, the human fingerprint, which is independent of rotation and scaling, is recognized. The multiple classification technique, based on wavelet and fractal analysis, is used. It is shown that systematic incorporation of decision from various classifiers leads to a better decision rather than simply fusing them. Multiple classifiers can serve as a means of enhancing the performance of pattern recognition problems. Multiple classifier system design involves the problem of classifier fusion. This paper deals with multi-classifier systems in which each classifier uses its own representation of the input pattern, based on features collected from multiple sources. The multiple feature sources considered here are multi-fractals, wavelets and fast Fourier transforms coefficients. A clustering algorithm is used to observe the efficacy of the feature sources. The multiple sources were graded according to their effectiveness of providing more non-overlapping clusters for different groups into which the samples are to be separated. This approach first considers the best source for the feature parameters. If this feature classifies the test sample into more than one cluster, then the feature next to the best is summoned to finish up the remaining part of the classification process. The continuation of this process along with the judicious selection of classifiers succeeds in identifying a single cluster for the test sample. The results obtained after the experiments on a set of fingerprint images shows that this novel technique can go a long way in avoiding ambiguity and thus limiting the need for use of soft-computing tools for making decisions. Our method provides a hard, concrete and accurate solution to pattern recognition problems employing multiple classifiers.