Automatic Thinning Selection Based on 3D Grape Bunch Reconstruction

Berry thinning in table grape production is a crucial operation that reduces the number of grapes before growth to promote berry enlargement and to improve bunch shape. However, the complex judgment criteria and need for quick decision-making make it extremely time-consuming for inexperienced workers to master this skill.To assist farmers in learning berry thinning expertise, this study proposes a system that performs 3D reconstruction of grape bunches and automatically identifies optimal berry selection for thinning. Conventional depth sensor-based and image-based 3D reconstruction methods have proven impractical in field environments due to lighting variations and the textureless surface characteristics of grape berries. Therefore, we propose a novel approach combining 3D Gaussian Splatting with Random Sample Consensus (RANSAC)-based sphere detection. Furthermore, combinatorial optimization using genetic algorithms is applied to the estimated berry positions, automatically determining which berries to thin by maximizing inter-berry distances.Evaluation of sphere position estimation accuracy using a grape bunch model achieved high precision with an average error of 4.03 mm. Berry detection experiments using 4 frozen Shine Muscat bunches demonstrated an average recall of 92.2\% with precision ranging from 98.4% to 100.0%, and the genetic algorithm successfully selected berries for removal from dense regions.We also validated the practical applicability in a real vineyard environment.In the experiment, 38 out of 39 berries were accurately detected, and appropriate thinning selections were made, confirming the effectiveness of the proposed method for field applications.