Reproduce “YOLO26-RipeLoc Lite: A lightweight architecture for tomato ripeness detection and picking point localization in greenhouse robotic harvesting” on OpenResearch

About this paper

Rajmeet Singh, Manveen Kaur, Shahpour Alirezaee, Irfan Hussain

Researchers from Khalifa University and the University of Windsor introduced YOLO26-RipeLoc Lite, a lightweight deep learning architecture that simultaneously performs real-time tomato ripeness detection, classification, and precise picking-point localization for greenhouse robotic harvesting. The model achieves 95.2% precision for ripe tomatoes with 2.38M parameters and localizes picking points with a 3.80 mm physical error at a 500 mm working distance, outperforming other nano-scale YOLO models in precision and efficiency.

Key insights

• Domain-specific augmentations, like greenhouse-aware HSV, and a 3-phase progressive unfreezing training strategy are crucial for improving model robustness and performance on small agricultural datasets under variable greenhouse conditions.
• The integration of a Ripeness-Aware Attention Module (RAAM) with a learnable ripeness bias effectively enhances the model's ability to discriminate between ripe and unripe tomatoes based on subtle color and texture cues.
• Designing a lightweight neck (LFPN) and a compact detection head (CDH) with shared weights and an integrated center-point regression branch enables efficient, single-stage real-time inference on resource-constrained edge devices without sacrificing accuracy.

Summary by alphaXiv.