A Peduncle-Aware Instance Segmentation and Geometric Localization Framework for Cluster Tomato Harvesting

Reliable segmentation of cherry tomato peduncles is difficult in greenhouse harvesting because peduncles are slender, occupy few pixels, resemble shoots and petioles, and are frequently occluded by clustered fruit. This study presents a peduncle-aware instance-segmentation and geometric-consistency framework for cluster-tomato harvesting perception. From the public 2022 Shanxi Agricultural Valley Tomato Town greenhouse cluster-tomato dataset, 1383 images with visible peduncles were selected and re-annotated in Labelme with polygon masks for fruit-cluster regions and peduncles. The framework combines boundary-weighted mask optimization, Coordinate Attention, and Soft-NMS. Evaluation on 208 held-out images showed a class-averaged mask AP50 of 0.9606 under the high-recall configuration; fruit-cluster and peduncle AP50 values were 0.9900 and 0.9313, respectively. At a score threshold of 0.50, overall precision, recall, F1 score, and instance-level Jaccard score were 0.8440, 0.9495, 0.8937, and 0.8078. In the 2D mask-geometric consistency evaluation, centroid-derived proxy points yielded matched-pair mean errors of 8.49 pixels for the peduncle and 5.94 pixels for the fruit-cluster region. These values quantify consistency between predicted-mask-derived and ground-truth-mask-derived points and do not represent physical robotic positioning accuracy.