To address the pollination strategy issues caused by the morphological differences of kiwifruit flowers at different blooming stages and the limited computing power of edge devices, this study proposes a lightweight kiwifruit flower detection method based on the improved YOLOv11n. The method accurately identifies flowers with different opening degrees, and compared with the baseline model, the model performs better on edge computing devices. Specifically, the StarNet lightweight backbone was employed to replace the feature extraction network of YOLOv11n, thereby reducing model complexity; an RGSE module integrating multibranch convolution and attention mechanisms was designed to enhance multiscale feature aggregation; the detection head was redesigned into a DetectEfficient structure to improve inference efficiency; and the loss function was replaced with MPDIoU. Experimental results demonstrate that the improved model, YOLOv11n‐SREM, achieves a precision of 93. 0%, recall of 92. 6%, and mean average precision (mAP) of 95. 3%. The number of parameters, model weight, and computational cost are reduced to 1. 5 M, 3. 3 M, and 3. 4 GFLOPs, representing decreases of 42. 3%, 41. 1%, and 46. 0% compared to the baseline YOLOv11n, respectively. The model achieves 303. 1 FPS, with 24 ms latency and 90. 1% detection accuracy on an edge device. These findings provide an effective technical solution for the lightweight deployment of automated kiwifruit pollination systems.
Zhang et al. (Sat,) studied this question.