Subpixel detection of peanut particles in wheat flour using near infrared hyperspectral imaging

Hyperspectral imaging in near-infrared region (NIR) is a powerful tool for characterization and detection in food industry. In particular, the scan of powders is a subject of interest for adulteration evaluation. However, such samples involve intimate mixture hence complex non linear effect in the reflectance signal. In this study, Adaptive Matched Subspace Detector (AMSD) is implemented for detecting peanut flour adulteration in wheat flour. The method consists of a hypothesis test based on the linear mixing model. This is compared with a non supervised technique based on Principal Component Analysis rejection method. Results show that AMSD performs the best by detecting adulterated pixels in samples with global concentration from 20% down to 0.02%. A coefficient of determination of 0.90 is obtained between the number of detected pixels and the global concentration of samples. PCA rejection method shows relevant but insufficient results by detecting much fewer adulterated pixels than AMSD. This study shows that the implementation of AMSD is successful and more efficient than rejection method based on the inner product variability.