TSP-Net: A Structure-Aware and Geometry-Constrained Network for Cherry-Tomato Truss Detection and Picking-Point Localization in Greenhouse Harvesting

During cherry-tomato-bunch harvesting, clustered fruits, thin stems, and ambiguous fruit–stem junctions cause missed detections and errors in picking-point localization. To address these challenges, we propose TSP-Net, a structure-aware and geometry-constrained detection network built on YOLOv11n, together with an ROI heat-map regression module for picking-point localization. TSP-Net integrates the Truss-aware Multi-scale Ghost Cross Stage Partial (TMG-CSP) module into the backbone to strengthen structural features of fruit edges, small stems, and clusters. It introduces the Truss-oriented Axial-Local Attention (TALA) module during feature fusion to capture fruit-arrangement direction and local occlusion boundaries. We also design Physics-informed Truss Consistency (PTC) Loss to regularize detections by enforcing consistency with fruit-cluster morphology through constraints on center distribution, inter-fruit spacing, and scale continuity. For picking-point localization, local ROIs are extracted from detection outputs, Gaussian heat maps are predicted, and continuous coordinates are decoded using soft-argmax. In the controlled evaluation, TSP-Net attains precision 81.81%, recall 73.21%, mAP@50 79.53%, and mAP@50:95 64.83%, which exceed the corresponding YOLOv11n results by 0.83, 0.32, 1.29, and 1.03 percentage points, respectively. The model size is reduced from 5.23 MB to 5.08 MB, the parameter count falls from 2.59 M to 2.49 M, and computational cost declines from 6.4 to 6.0 GFLOPs. On the common valid-ROI test set, the proposed heat-map localization yields a mean localization error of 47.2 px, a median error of 24.2 px, and PCK@20, PCK@50, and PCK@100 of 45.1%, 68.4%, and 85.5%, respectively. Under identical ROI conditions, this heat-map method reduces the mean localization error by 11.3 px and raises PCK@20 by 14.9 percentage points compared with direct coordinate regression. In summary, the proposed approach enhances cherry-tomato-bunch detection and 2D picking-point candidate localization while meeting lightweight constraints, offering a structure-aware visual perception solution for greenhouse cherry-tomato harvesting.