Abstract: The size of ‘Red Globe’ grape bunches is one of the most important quality indicators. A method for size grading of ‘Red Globe’ grape bunches was proposed based on machine vision in this paper. The Red-Green-Blue (RGB) and near infrared (NIR) images of grapes were collected simultaneously with a two-channel camera. Subsequently, the information about green stalks was removed by the normalized super green method from the RGB images, and the local maxima of brightness based on morphological reconstruction was used to identify and locate the grapes in the NIR images. The contours of the fruits were effectively cut and the interference arcs in the edge contour were removed by the gradient segmentation method. Then, the size of the grapes was obtained by random least squares ellipse detection algorithm. A total of 42 bunches of grapes were graded by this method, and 38 of them were graded correctly with an accuracy rate of 90.48%. The accuracy rate is high enough to provide technical support for commercial grape grading.

Key words: ‘Red Globe’ grape, size, normalized super green method, morphological reconstruction, gradient segmentation, random least squares