Abstract:  To achieve realtime detection on resourcelimited embedded devices, a lightweight detection model （YOLOv5s_SGSC） for kiwifruit flowering period based on the improved YOLOv5s is proposed. Based on the YOLOv5s model, the ShuffleNetv2 and Ghost convolution are used to replace traditional convolutions in the backbone feature extraction network and the neck network, respectively. A convolutional attention module （CBAM） is embedded to improve the networks ability to extract features from kiwifruit flowers. The experimental results show that the precision and recall of the improved model are 89.9% and 89.7%, with a mAP value of 94.5%, which is a 0.3 percentage points increase compared with the model before the improvement. The model size is 3.9MB, only 27.7% of the original YOLOv5s model. The realtime detection speed on embedded devices is 11.8fps, which is 59.8% faster than the original YOLOv5s model. The model is deployed on embedded devices for field experiments. The experimental results show that the improved model correctly identifies kiwifruit flowering at distances of 20-60 cm with an accuracy of more than 85%, with a realtime detection frame rate above 10fps. Therefore, this study can effectively classify the flowering period of kiwifruit flowers, contributing to the development and application of pollination robots.

Key words: kiwifruit flowers, flowering identification, embedded devices, YOLOv5s algorithm, lightweight

摘要： 为在有限的嵌入式设备资源下达到实时检测要求，提出一种基于改进YOLOv5s的猕猴桃花期轻量化检测模型（YOLOv5s_SGSC）。在YOLOv5s模型基础上，使用ShuffleNetv2和幻影卷积分别替换主干特征提取网络和颈网络的传统卷积，嵌入卷积注意力模块（CBAM）提高网络对猕猴桃花朵的特征提取能力。结果表明，改进后模型的精确率和召回率为89.9%和89.7%；mAP值为94.5%，较改进前提高0.3%。模型体积为3.9MB，为原YOLOv5s模型的27.7%，在嵌入式设备实时检测速度为11.8fps，比原YOLOv5s模型快59.8%。将模型部署到嵌入式设备进行实地试验，改进后模型对距离镜头20~60cm的猕猴桃花朵花期正确识别率达到85%以上，实时检测帧率在10fps以上。可实现对猕猴桃花朵的花期分类，有助于推动授粉机器人的研发与应用。

关键词: 猕猴桃花朵, 花期识别, 嵌入式设备, YOLOv5s算法, 轻量化