Artificial Medjool Date Fruit Bunch Image Synthesis: Towards Thinning Automation

Highlights Medjool date fruit bunches can be modeled in 3D based on structural decomposition and the use of Bezier curves. The 3D model can be used for generating artificial image datasets of Medjool fruit bunches. The annotated image datasets can be used to develop robust algorithms for robotic Medjool date thinning. Algorithms for determining the required thinning length are a prerequisite for Medjool date thinning automation. Abstract. Medjool is a premium date cultivar, and the market demands high-quality fruits, for which specific horticultural practices, including timely and efficient fruitlet thinning, are required. Currently, thinning the fruitlets is one of the most labor-intensive tasks in the Medjool cultivation cycle, and there is a need to develop methods for automating the thinning process. An algorithm determining the required thinning is a prerequisite for advancing toward thinning automation. An annotated Medjool fruit bunch image dataset is necessary for developing such an algorithm using state-of-the-art machine learning methods. Acquiring such a dataset is difficult and costly. The difficulty can be alleviated by using synthetic images. However, current methods for generating synthetic plant images are unsuitable for Medjool dates due to their geometrical features. The current work suggests a method for generating artificial images of Medjool fruit bunches from a 3D model based on structural decomposition into plant parts and the use of Bezier curves. Nineteen model variables and their distributions were defined for fruit bunch model generation. The models and synthetic images generated based on the models were verified by two plant physiologists who are experts in Medjool date cultivation. Fruit-bunch features were extracted from the generated images and used for learning the required remaining length of the spikelets after thinning using kernel estimation. The estimation was tested for images of two whorl-period combinations (Top-Early and Middle-Middle). The average scaled absolute estimation errors for both periods were very low (less than 1%).