Feature Extraction Methods Based On Complex Wavelet Transform

With the development of the Internet and multimedia technologies, the volume of various data is dramatically increasing and data analysis and knowledge acquisition become more and more difficult. Data classification is one of the most important data analysis methods and feature extraction is the basis of data classification. It is an advisable technique to transform the raw data and extract features in the transformed domain using signal processing methods.Wavelet transform has been applied to data feature extraction and good results have been achieved. However, wavelet transform has spectrum aliasing that causes serious shift variance, especially in processing of two-dimensional image signals. Therefore it will affect the ability of wavelet coefficients to characterize the signal and thus will reduce the efficiency of classification. There are obvious distinctions in feature extraction methods based on signal transformation in terms of transformed coefficients selection and features construction. In view of these facts, this thesis selects complex wavelet transform to process data and designs novel feature extraction algorithms in the complex wavelet domain, and experimental results show that these feature extraction algorithms can improve correct classification rate.The research work can mainly be concluded as follows:First of all, we design a feature extraction method for one-dimensional signals in complex wavelet domain based on the central limit theorem and Lyapunov theorem. The method can extract more accurate characteristics of one-dimensional signals, and it can be applied into feature extraction for many types of one-dimensional data.Secondly, we complete classification experiments of one-dimensional sequence databases based on the above method. Experiments have been carried out on protein sequences, voice signal sequences and brainwave signal sequences respectively. The results show that this method can achieve good classification accuracy in these three kinds of one-dimensional data and hence verify the applicability of the method.Thirdly, we propose a method of texture image feature extraction in complex wavelet domain based on association rules mining. The statistical parameters of mined association rules are structured as image features and applied in image classification. This method can reveal the hidden characteristics in the image micro structures and experimental results show that this method can lead to higher correct classification rate.