Terahertz signal analysis and substance identification via Zernike moments

Terahertz (THz) spectra contain chemical information, along with noise and variable backgrounds. Measurement environmental changes and spectral signal differences caused by changes in the sample state can degrade the accuracy of the calibration models. This problem obviously hinders practical applications of THz spectroscopy. To tackle this problem, a three-dimensional spectrum was first self-constructed and converted into an intensity image. Zernike moments with inherently invariant properties were then used to describe the THz intensity image and extract the invariant features for further analysis. Considering the reconstruction error and computational cost, the highest order of Zernike moments and the most effective moments were selected and applied to multi-classifiers including support vector machines, naive Bayes, and regularized linear discriminant analysis. Experiments used a THz dataset collected from four chemical substances (melamine, tartaric acid, lactose, and glucose) at five thicknesses (1.0 mm, 1.5 mm, 2.0 mm, 2.5 mm, and 3.0 mm). The results confirmed the effectiveness of the proposed approach. The obtained results show that compare to traditional absorption spectrum features, Zernike moment features are less sensitive to spectral variations caused by changes in sample status. They have better feature representation ability with lower feature vector dimensions. This suggests that they can be integrated into the design of systems for THz spectral classification to increase the robustness and generalization capability of the classifier.