Abstract:

In the present study, based on the RGB images acquired using a 3-CCD high-precision area array camera for
several varieties of walnuts in Xinjiang, we designed a self-adaptive double-threshold Otsu method which can rapidly and
accurately segment the defective regions and transform 20 initial features including geometry and texture and other features
to a 9-demensional set of eigenvectors. Using the set of eigenvectors as input, 15 recognition models were established based
on Bayesian network, BP neural network (BPNN) and support vector machine (SVM), and their adaptability as well as
identification performance and mean recognition time for 3 defects (crack, damage, and black spot) were compared. The
results revealed that the SVM model based on radial basis function (RBF), showing a mean recognition time at the order
of magnitude of 10-4 s, provided the best results, giving average test recognition accuracy of 93.06% for crack, 88.31% for
damage, and 89.27% for black spot and total recognition rate of 90.21% for the 3 external defects. These results can provide
useful data for on-line determination and classification of walnut detects and on-line quality identification of other nuts.

Key words: walnuts, machine vision, external defects, support vector machine, recognition