Answers are grounded in this paper and cite their sources.
Tomato is a significant economic crop cultivated extensively worldwide. Rapid and accurate localization of picking points is fundamental to automating the tomato harvesting process. To address this, we develop a lightweight dual-task network built upon the YOLOv10 (You Only Look Once) architecture. The backbone integrates the ADown module for efficient downsampling, and utilizes a Dysample-based cross-scale feature fusion module (CCFM) in the detection neck, while the segmentation neck employs Simulated Annealing Adaptive Concatenation (SAAC) for feature fusion. Furthermore, the conditionally restricted Hungarian algorithm is employed to match ripe tomatoes with their corresponding corollas. Two localization techniques are proposed: local skeleton line fitting and boundary intersection screening. In local skeleton line fitting, the corolla detection box is expanded, and the stem mask within this region is used to extract the skeleton line and find the intersection point. Boundary intersection screening determines whether to directly compute the midpoint of intersecting lines or use skeleton line fitting, based on the number of intersections between the stem mask and the corolla box. The detection Precision and Recall of the proposed model are 0.970 and 0.983, respectively, while the IoU (Intersection over Union) and MIoU (Mean Intersection over Union) for segmentation are 0.962 and 0.980. The localization precision for picking points is 0.922, with a recognition rate of 0.988. Achieving an average processing time of only 55.89 ms per image, the proposed artificial intelligence application exhibits excellent efficiency and accuracy in real-world conditions, offering a promising foundation for further development of automated tomato harvesting systems.
No takes yet. Share an insight, caveat, or question.
An et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: