Intelligent Thinning Decision for Strawberry Quality Optimization: A Lightweight Peduncle-Fruit Topological...

Fruit thinning concentrates nutrients and serves as an essential procedure for cultivating high-quality strawberries with exceptional palatability. It is established that controlling the number of fruits per peduncle to three to five significantly increases the proportion of large fruits and enhances sweetness while effectively reducing the risks of disease infection. Consequently, pesticide application is minimized, thereby ensuring food safety and improving overall product quality. Although precise execution baselines are a prerequisite for automated thinning, current research is largely restricted to isolated fruit recognition, leaving the direct detection of peduncle-fruit topological associations under complex occlusions unaddressed. To address this gap, a lightweight, three-stage ("coarse-to-fine") detection method tailored for automated thinning is proposed. The pipeline comprises global localization, local cropping with background suppression, and fine secondary inference, which effectively mitigates environmental interference. To support this investigation, a dedicated dataset comprising 1131 high-quality images was constructed. For efficient edge deployment, YOLOv11n was adopted as the baseline architecture, integrated with structural Re-parameterized Convolution (RepConv), Coordinate Attention (CoordAtt), and a dynamic re-weighting loss. This configuration ensures robust feature extraction of slender peduncles with an extremely low parameter overhead. The experimental results demonstrate that with only 2.77 M parameters, the proposed model achieves an mAP@0.5 of 75.73% and an F1-score of 73.85%. Notably, the average false positive (FP) detections per image in complex scenarios were significantly reduced from 1.53 to 0.23. Following deployment on a Jetson Orin Nano Super edge device utilizing TensorRT and FP16 quantization, the end-to-end system inference speed stabilized at 15.58 frames per second (FPS) with near-lossless precision. Ultimately, this method provides a reliable technical foundation for automated thinning decisions, facilitates sustainable greenhouse management, and secures the supply of high-quality food from the source.