ABSTRACT
In the field of food testing, variety, brand, origin, and adulteration are four important testing projects. In this chapter, we propose a novel negentropy sorted kernel independent component analysis (k-ICA) method as a feature extraction method for Fourier-transform infrared (FTIR) spectroscopy. Since the independent components (ICs) obtained by ICA is random, it needs some criteria to sort these ICs. Here, we use negentropy as a criterion to measure the non-Gaussian of ICs in order to separate the IC from maximum negentropy first. Then we use a kernel support vector machine (k-SVM) as the classifier to distinguish different foods. We use four datasets to comprehensively investigate four kinds of problems (variety, brand, origin, and adulteration). The experimental result indicated that the k-ICA presents a superior performance than traditional method such as plusLrR, Fisher, and principal component analysis (PCA), also it is better than using original wavelength. The k-SVM model has the best performance, and it is better than back propagation-artificial neural network (BP-ANN) and partial least squares (PLS). The improved double kernel method (k-ICA and k-SVM) can detect food’s variety, brand, origin, and adulteration simultaneously, and the recognition performance is steady, high, and efficient, and the recognition program works steadily efficiently. This conclusion has positive significance for food detection.
