Real-time Fruit Detection Using Deep Neural Networks

Barna Keresztes (1) , Florent Abdelghafour (1) , Dimby Randriamanga (1) , Jean-Pierre da Costa (1) , Christian Germain (1)

1 IMS - Laboratoire de l'intégration, du matériau au système

Proximal imaging using tractor-mounted cameras is a simple and cost-effective method to acquire large quantities of data in orchards and vineyards. It can be used for the monitoring of vegetation and for the management of field operations such as the guidance of smart spraying systems for instance. One of the most prolific research subjects in arboriculture is fruit detection during the growing season. Estimations of fruit-load can be used for early yield assessments and for the monitoring of harvest and thinning. In addition, the visual aspects of fruits enable to appraise their growth and ripening status. This paper proposes a new approach for real-time fruit detection, combining a fast geometrical pre-processing whose output feeds a deep neural network (DNN) classifier. The first step is a radial Hough-like operator, which aims at identifying quickly the regions of interest, restricting the use of the DNNs to the most probably genuine candidates. The proposed method is generic enough to be applied on most near-spherical fruits. It was tested in two contexts: grapes and apples, with different varieties and phenological stages. In both cases the proposed method provided promising results. Correlation coefficients with manual counting and real harvest loads are up to 0.96 for grapes and up to 0.85 for apples.

Mots clés: Proximal sensing; image analysis; fruit detection; deep learning; radial Hough transform

Domaines: Traitement du signal et de l'image [eess.SP]

Barna Keresztes, Florent Abdelghafour, Dimby Randriamanga, Jean-Pierre da Costa, Christian Germain. Real-time Fruit Detection Using Deep Neural Networks. 14th International Conference on Precision Agriculture, 2018, Montréal, Canada. ⟨hal-02518559⟩