Support vector machines combined with wavelet-based feature extraction for identification of drugs hidden in anthropomorphic phantom

Abstract A new recognition method of Support Vector Machines (SVMs) combined with wavelet-based feature extraction is proposed for identifying drugs hidden in human body. Preliminary data sets of eight kinds of samples are acquired by a home-built instrument using energy-dispersive X-ray diffraction (EDXRD) technology in a short detection time. Small sample size, poor signal-to-noise ratio (SNR) and high dimension of data make drugs identification a challenging problem. In this paper, the potential effective method solves the problem well. The spectral signal with poor SNR is obtained and processed with wavelet for feature extraction and then the wavelet coefficients are used as the inputs of SVMs. A multi-classifier of SVMs based on binary tree architecture (SVMs-BAT) is trained. The method of SVMs-BAT combined with wavelet-based feature extraction (WSVMs-BAT) is firstly compared with two methods: one is single SVMs-BAT which uses original data as inputs without preprocessing, the other is SVMs-BAT combined with feature extraction based on principal component analysis (PCA-SVMs-BAT). The high identification accuracy of WSVMs-BAT indicates that the method of feature extraction using wavelet can effectively represent the original data better. Then the recognition result of the proposed method is also compared with artificial neural network (ANN) and K-nearest neighbor (KNN) methods. Our findings show that the proposed method combined with EDXRD technology provides a good access to achieve the aim of automatic identification of illicit drugs.