End-to-end lightweight berry number prediction for supporting table grape cultivation

• DOI: 10.1016/j.compag.2023.108203
• Publication Date: 2023-09-23T10:00:47Z
• Authors: Yan San Woo; Prawit Buayai; Hiromitsu Nishizaki; Koji Makino; Latifah Munirah K...; Xiaoyang Mao
• Keywords: Minimum bounding box; Table grape; Feature (linguistics); Bounding overwatch; Machine Vision

The advent of smart agriculture has revolutionized and streamlined various manual tasks in grape cultivation, one which is berry thinning. This task necessitates experienced farmers to selectively remove a specific number berries from the working bunch, as guided by remaining bunch. In response, this paper introduces novel real-time edge computing application that automates process counting bunch using single 2D image. proposed employs YOLOv5-based object detection techniques (Jocher, 2021) distinguish each visible slightly occluded contained therein. key contribution accurately predict whole bunches including those not image harnessing output devise features based solely on bounding box information. addition, feature set optimized employing wrapper selection method (Kohavi & John, 1997), consideration limitations devices. eight selected yield mean absolute error (MAE) 2.60 berries, tested dataset 26,230 images. Only slight increase over initial 19-feature set, achieved an MAE 2.42 berries. Furthermore, approach been successfully implemented Android smartphone, Sony Xperia 1 III, without need for internet connection. overall computation time per stands at average 0.333 s, confirming its potential real-world application.