Lightweight Multi-Occlusion Pear Detection via Multi-Auxiliary Domain Transfer Learning for Robotic Harvesting

Accurate detection of occluded pears is vital for selective robotic picking. However, existing methods face critical challenges: an inadequate trade-off between lightweight model design and detection accuracy, which restricts deployment on resource-constrained robotic platforms; a domain shift issue in transfer learning, resulting in long training times and wasted computational resources; and oversimplified single-category classification that misidentifies occluded fruits, causing picking failures and hardware damage. We propose a lightweight detector empowered by multi-auxiliary domain transfer learning (MADTL) for accurate multicategory pear detection. Specifically, built upon YOLOv8, the proposed detector optimizes the backbone and neck architectures by integrating advanced modules to enhance feature extraction and fusion efficiency. Crucially, the proposed MADTL strategy introduces apple and orange datasets to bridge the source-target domain gap, significantly accelerating convergence. Benchmarked against YOLOv8s, our detector reduces model size by 62.4% and floating-point operations by 53.7%. Notably, MADTL accelerates convergence by 75% while boosting accuracy. Field deployment achieves real-time inference at 47.39 ms per image. These improvements enable real-time deployment on resource-constrained edge devices while maintaining high detection accuracy, providing essential support for selective harvesting to minimize picking failures and enhance operational efficiency.