Research On Some Feature Extraction Methods

Feature extraction is considered as one of the most fundamental research contents. It caneffectively alleviate ‘curse of dimensionality’ in the field of pattern recognition. Moreover, itplays an importance role for recognition performance. Feature extraction has been widelyapplied to the fields of biometrics recognition, information processing, and text categorization.Till now, there are lots of feature extraction methods. Unfortunately, there exist manyshortcomings in the existing approaches, such as being sensitive to noise, existing small sizesample problems, and existing redundant information in the extracted features.In this dissertation, some researches have been made on the two-dimensional featureextraction methods and the feature extraction method for one-class classification. The foresaidshortcomings have been overcome and the main contributions of the dissertation aresummarized as follows.1. The traditional2DLPP is based on L2norm and sensitive to outliers. In thisdissertation, the L1-Norm based2DLPP (2DLPP-L1) has been proposed.2DLPP-L1utilizes an iterative algorithm to reduce the dimensions of the given data set.Moreover, it can avoid eigenvalue decomposition. In comparison with the L2-normbased2DLPP, the proposed L1-norm based2DLPP is more robust against noises.2.2DLDA essentially works solely in the row-direction of images. Therefore, thefeatures extracted by2DLDA may contain redundant information. To solve thisproblem, the two-stage dimensionality reduction approach based on2DLDA andfuzzy rough sets technique (2DLDAFRS) has been proposed. In the first stage,2DLDA is performed to extract features. In the second stage, the fuzzy rough setstechnique according to the frequency of the features is conducted to get rid of theredundant features extracted in the first stage.3. The parameter setting of a certain one-class classifier, i.e. SVDD is very difficult.Moreover, there exist redundant features in the give data set. In this dissertation, thesimulated annealing approach for feature extraction and parameter selection ofSVDD, namely, SA-SVDD has been proposed. During the procedure of simulatedannealing, the optimal kernel parameter, trade-off parameters, and number ofextracted features are automatically selected.Experimental results demonstrate that the proposed methods can improve the extraction efficiency and the performance of their corresponding traditional methods.