Nondestructive Identification of Star Anise and Shikimmi by Visible/Near Infrared Hyperspectral Images

Based on hyperspectral imaging technique, an identification method of star anise and its counterfeit shikimmi was proposed. The hyperspectral data in the range of 400~1 000 nm were collected and analyzed. Firstly, according to the different spectral characteristics of samples and background pixels, images at 850 nm and 450 nm were selected and subtracted, and background information was removed by threshold method. Linear stretching method was further used to remove shadow noise pixels due to sample height. Combined with the region labeling method of binary image, the automatic extraction of average spectral data from sample hyperspectral data was realized. Then based on average spectral data, four optimal wavelengths were selected by successive projections algorithm (SPA), i.e., 533 nm, 617 nm, 665 nm and 807 nm. Based on the spectral data at the optimal wavelength, a partial least square discrimination analysis (PLSDA) model was established. The classification accuracy of star anise and shikimmi was 98.4%. Using the established multi-spectral model to predict the external validation set data, the overall classification accuracy was 97.9%, and the visualization results were good. Finally, the conventional image processing technology was also used to process the same external verification set data, and the results and advantages of the two methods were compared. The results showed that the multispectral analysis method based on hyperspectral imaging technique was simple, efficient and easy to realize dynamic on-line or portable detection applications. The proposed method can provide a theoretical basis for the development of on-line or portable detection instruments. © 2019, Chinese Society of Agricultural Machinery. All right reserved.