Çapraz Sınıflandırma: Teorik Çerçeve Ve Uygulamalar

Abstract

Pattern identification algorithms can be exploited for single-class or multi-class identification problems. In conventional applications, identification algorithms are based on functions of measured features. However in some cases, we already have an identification algorithm but the measured features are different from the expected features. In such cases, we can first estimate the needed features from the measured ones and then use the identification algorithm with the estimated features. This work proposes a framework for such Cross Classification problems. This study can be grouped under four different headings. These can be summarized as follows; 1-distance measures, 2-Linear Minimum Mean Square Error (LMMSE) estimator based cross classification, 3-noisy feature modelling for cross classification and 4-Canonical Correlation Analysis base cross classification method. Distance measure is one of the main instruments of statistical pattern recognition and classification problems. It is used to measure the separation of the two objects. Bhattacharyya Distance is one of the statistical distance measures used for class separability problems. This distance measure is utilised in this study for investigation of different cross classification approaches, because of its usability in limit values for classification. LMMSE based method is one of the most basic methods used for the cross-classification. An LMMSE predictor based on the direct prediction of unmeasurable features principle is defined in this study and it is applied to the cross classification problems. The performance of this approach is investigated in detail. The noisy feature model is based on modelling of the prediction errors as noise on unmeasurable features. This model enables the analytic study of the classification performances after prediction. The correlation between Bhattacharyya Distance and classification performance iii is analysed using noisy feature model. The utilisation of CCA methodology for cross classi- fication is investigated and a method is proposed in this work. Moreover, the advantages of this method for feature selection and reduction is presented. The relation between LMMSE based and CCA based cross classification is studied. The CCA based classification is presented to perform similar to LMMSE method when all features are employed. In this study we have also shown the applications of different crossclassification problems using synthetic and real data sets. The proposed methods are applied to face recognition from sketches problem and the results are presented with comparisons to other methods in literature. In addition, the presented cross classification method is used for the face recognition from facial pictures captured from different distances problem and the results are discussed