Detection of Bruise on Pear by Hyperspectral Imaging Sensor with Different Classification Algorithms

A hyperspectral imaging sensor system was developed for the detection of bruises on pears, for these bruises were difficult to be detected by traditional computer vision technique. Hyperspectral imaging sensor technique is susceptible to the effects of uneven illumination due to a spherical object of pear. The data of hyperspectral image is a 3-dimension cube, which contains a huge amount of information. So it requires a suitable algorithm to extract some useful information from the 3-dimension data cube. In this work, Principal Component Analysis (PCA) was firstly used to extract some useful information, then several other classification algorithms were used comparatively to process the 3-dimension data cube. These classification algorithms were Maximum Likelihood Classification (MLC), Euclidean Distance Classification (EDC), Mahalanobis Distance Classification (MDC) and Spectral Angle Mapper (SAM), respectively. Results show that MDC and SAM have well performance, with detection accuracy of 93.8% and 95.0% respectively. Compared with the other classification algorithms, MDC and SAM can overcome the effects of uneven illumination in detecting bruise of pear by hyperspectral imaging sensor technique. This work demonstrates that it is feasible to detect the bruised region on the surface of pear by hyperspectral imaging sensor technique combined with MDC and SAM.