Near infrared hyperspectral imaging technology was adopted in the paper to determine the variety of raisins produced in Xinjiang Uygur Autonomous Region, China. There are 2 varieties of raisins taking part in the research and the wavelengths of the hyperspectral images are from 900nm to 1700nm. A novel waveform resolving method was proposed in the paper to reduce the hyperspectral data and extract features. The waveform resolving method compresses the original hyperspectral data for one pixel into 5 Amplitudes, 5 frequencies and 5 phases, 15 feature values in all. Neural network was established with three layers, 8 neurons for first layer, 3 neurons for hidden layer and 1 neuron for output layer based on the 15 features to determine the variety of raisins. The accuracies of the model are higher than the accuracies of model of traditional PCA feature extracting method combined with neural network. The result indicates that the proposed waveform resolving feature extracting method combined with hyperspectral imaging technology is an efficient method to determine variety of raisin. The 2 varieties of raisins spectra information of wavelength were acquired by hyperspectral image system and the varieties of raisins were classified with ideal performance by the proposed SFEWR and neuron network method. After fast Fourier transform, the 226 original wavelengths were compressed to 15 features which can express the most information of the spectrum. After neuron network model established based on the 15 features, the raisins can be classified according to the varieties and the performance of the proposed SFEWR combined with neuron network can be expressed by 93.93% of sensitivity, 93.93% of precision and 98.98% of specificity which are better than PCA with neuron network. The detected results being expressed by images show that the proposed method is very successful in classifying raisins while there are some suggestions should be considered in the future researches:

1 Consider some ways to overcome the problems which are due to different distances the light reflecting from sample to lens.

2 Separating 8 varieties at the same time leads to very narrow distances among them, and also the range of error permissible is very narrow. Thus bisection method could be taken into account that is the 4 varieties of 8 with similar features are considered as the first class and the other 4 as the second, then the first class separated into 2 classes and so do the second class, and so forth until 8 varieties are separated completely.