RrT-YOLO: a scale-adaptive framework for Rosa roxburghii Tratt maturity detection in complex mountainous environments
Abstract
To address the challenges of dense fruit distribution, severe occlusion, substantial scale variation, and low visual contrast between partially ripe fruits and background foliage in Rosa roxburghii Tratt (RrT) orchards, we propose RrT-You Only Look Once (YOLO), a detection framework developed upon the YOLO architecture through coordinated optimization of feature representation, scale adaptation, and object localization. Specifically, the Scale-Adaptive Hybrid Shared Sparse Mixture of Experts module enhances fine-grained texture and boundary representation through shared feature transformation, input-conditioned sparse expert routing, and multi-scale feature fusion, while the Shared Batch-normalized Decoupled Head improves multi-scale prediction by combining scale-private batch normalization with shared spatial transformations and separate regression and classification projections, enabling parameter reuse while preserving scale-dependent feature responses and task-specific prediction. To further improve localization accuracy under dense and overlapping conditions, the Weighted Inner Point-wise Intersection over Union loss combines IoU overlap with angle-aware distance and shape constraints to stabilize bounding-box regression. Experiments on the self-constructed RrT dataset demonstrate that RrT-YOLO achieves a mean Average Precision at an Intersection over Union threshold of 0.50 (mAP50) of 72.8% and a Precision of 70.4% with 7.34 M parameters, while maintaining an inference speed of 158 frames per second, corresponding to a 6.3% improvement in mAP50 over the YOLO26 baseline. Cross-dataset evaluations on papaya, banana, and mango datasets further yield mAP50 values of 98.4%, 97.7%, and 95.3%, respectively, demonstrating favorable cross-category applicability across fruit species with distinct appearances, scales, and growth environments. These results demonstrate that RrT-YOLO provides a reliable visual perception framework for fruit maturity detection under complex orchard conditions while maintaining real-time inference capability, offering practical potential for intelligent harvesting and precision agriculture.