Identification of apple stem and calyx using unsupervised feature extraction

Neural networks and unsupervised feature extraction were used to classify apple images based on whether or not they included a stem or calyx end. In one experiment, the system successfully classified 98.4% of a test set consisting of 254 near-infrared images captured at 740 nm. The network was also tested on gray-level images captured with light in the visible range. In this test of 242 images, 5% were not classified correctly. In another classification test that included apple images with prominent defects, 28% were misclassified. However the majority of the errors occurred because major defects were mistaken for the stem or calyx. In a practical implementation, errors that could lead to loss in a sorting system would amount to only 0.05%.