Real-Time Multi-Attribute Quality Grading of Tabletop Strawberries Under Occlusions for Harvesting Robots

• Modular RGB–D pipeline for real-time multi-attribute grading of strawberries. • ResNet-101 multi-head classifier predicted maturity, disease, and deformity jointly. • CycleGAN restored strawberry shape and texture from partially occluded views. • QH and MADM strategies converted multi-attribute predictions to grading decisions. • Field tests validated end-to-end perception-to-placement pipeline in farms. Strawberry grading was critical for meeting market standards, yet manual post-harvest sorting remained labor-intensive, damage-prone, and inconsistent, while most vision-based methods targeted single attributes and lacked integrated in-field operation. We presented a real-time RGB–D grading framework for tabletop-cultivated strawberries in greenhouse environments. The system combined instance segmentation (YOLOv11–Seg), occlusion-aware completion via CycleGAN, and a three-branch multi-task network based on ResNet-101 to simultaneously infer maturity, disease status (including asymptomatic fruit), and deformity. Pose normalization and depth alignment supported calibrated volume-to-mass regression for non-destructive weight estimation. To improve robustness to rare morphology, 500 deformed-fruit samples were synthesized with a Gemini-based generative model to mitigate class imbalance. Built upon this perception stack, we implemented two grading strategies: a deterministic Quality Hierarchy (QH) for real-time robotic routing and a weighted multi-attribute decision-making (MADM) scheme for flexible batch evaluation. The multi-attribute classifier achieved 93.65% overall accuracy (disease 94.09%, maturity 94.07%, shape 94.42%) with an inference latency of 33.65 ms (29.7 FPS). Depth-assisted mass estimation yielded average errors of 8.11% for complete fruits and 10.47% under occlusion after completion; in field deployment on marketable fruits routed to size grading, it achieved MAE/RMSE of 1.85/2.20 g with R 2 = 0.9384 (MAPE 8.48%) and a three-bin size-grade accuracy of 90.91% (100/110). In single-target harvesting mode on an RTX 4060 GPU (640x480, 30 fps), end-to-end latency was 87.56 ms per harvested target in complete mode and 197.56 ms when completion was invoked (CycleGAN 110.00 ms/instance). These results demonstrated a practical, non-destructive, occlusion-aware, and deployable multi-attribute grading solution for intelligent strawberry harvesting in real greenhouse scenarios.