Colour segmentation of citrus fruits images for stem location

Abstract

After mechanical harvesting, Cirrus fruits with long stems may cause damage on adjacent fruits during transportation and storage. A first step to solve before measuring the stem and taking the decision of cutting it or not, is the orientation of fruits. For this purpose, an image analysis method to locate the stem insertion point based on colour segmentation is proposed and evaluated. Four classes were defined for segmentation: background, peel, stem-calyx and cut stem, and a classifier based on bayesian decision rules was developed assuming that the independent variables (RGB) followed a normal distribution function, and that the classes had equal covariance matrix. The results point out an excellent performance of the classifier for the first two classes, and satisfactory for the other two. The location algorithm used information of areas and centroids to estimate the coordinates of the stem insertion point. From a total of 86 images tested, it was correctly estimated a 90,3% of times. The method resulted to be efficient for fruits without leaves appended to the stem.