Contribution to fruit gradding by using numerical images analysis. Application to on-line gradding of Jonagold and Golden delicious apples.

This work aims to set up an automatic quality fruit grading method based on external characteristics. Special attention is drawn to various defects such as wounds, bruises, physiological diseases, fungi attack, etc. Two 3-CCD cameras mounted on a test rig were used for the image acquisition. Golden delicious apples are characterised by their uniform colour. This later was modelled by a multivariate Gaussian distribution and the defect detection was carried out by computing the Mahalanobis distance separating a pixel's colour and the mean colour of the fruit. For Jonagold apples, having a multimodal colour frequency distribution, the defect location was based on a non-parametric model of the fruit colour and on Bayes' theorem. In both cases, the development of an algorithm, taking into account local information, enhanced the segmentation precision. The calyx and stem ends, which appear as defects on the image, were detected by a pattern correlation technique. The segmented areas (poles, defects and over-segmentation zones) were characterised with shape, colour and texture descriptors. The fruit grading into four classes (Extra, A, B and cull) according to European standards is obtained by using a cluster analysis on the segmented regions. The results obtained are favourable and make it possible to envisage the transfer of developed algorithms onto a commercial sorting machine