基于改进YOLO 12s的甜瓜雌雄花识别与其边缘设备部署测试

Performance Evaluation of Muskmelon Flower Gender Identification Based on Improved YOLO 12s and Its Deployment on Edge Devices

摘要:

授粉是甜瓜种植的关键环节，机械授粉因其高效性成为重要的发展方向，雌雄花朵精准识别是机械授粉作业的重要保障。本研究提出了一种基于改进的YOLO 12s目标检测网络的甜瓜雌雄花朵识别方法并进行了嵌入式系统平台部署。通过将原模型的Backbone部分替换为ShuffleNetV2结构降低模型复杂度与计算量，以提升检测速度，使其适于边缘设备上的部署，并通过知识蒸馏进一步优化检测算法性能。对温室中拍摄得到的甜瓜雌雄花朵进行标注，利用改进的YOLO 12s网络进行模型训练，并在相同试验条件下，与当前主流算法进行对比，分别用精确率、召回率、平均精度均值(mean Average Precision, mAP)、模型内存占用量、检测速度5个指标对检测模型进行了性能评估。经测试，模型精确率为79.93%，召回率为87.65%，mAP为86.54%，模型内存占用量为4.10 MB，检测速度为153.46 f/s，与YOLO 12s、YOLO v10s、YOLO v8s、YOLO v5s、SSD和Faster R?CNN模型相比，其检测速度分别提升79.32%、75.11%、45.14%、38.23%、746.44%、1 168.26%，且在光照变化、部分遮挡及图像模糊等复杂场景下仍能实现精准检测。迁移部署于Jetson Xavier NX嵌入式平台，模型mAP为86.62%，检测速度为26.37 f/s，其中，检测速度分别比YOLO 12s和YOLO v5s高90.40%、15.51%，可满足实际授粉作业对精确度和实时性的需求。结果表明，经轻量化处理后的模型可准确、快速地实现甜瓜雌雄花识别，且在精度损失较小的前提下显著减小模型内存占用量，有利于网络在边缘设备上的部署，该研究结果可为甜瓜机械授粉提供技术支持。

Abstract:

Pollination is a critical stage in muskmelon cultivation, and mechanical pollination has emerged as a significant development direction due to its efficiency benefits. Accurate identification of female and male flowers is essential for the operation of pollination robots. For identifying melon female and male flowers, an improved YOLO 12s target detection network was applied, and the systematic deployment was implemented on edge device. By replacing the Backbone of the original model with the ShuffleNetV2 structure, the model?s complexity and computational load were reduced, and the purpose of improving detection speed was achieved. Additionally, the knowledge distillation was also employed to further improve the performance of the improved model. A total of 3 597 muskmelon images were annotated in the greenhouse environment and then sent to the improved YOLO 12s object detection network for model training. Under identical experimental conditions, a comprehensive comparison was conducted with current mainstream algorithms. The melon flower detection model was evaluated across five metrics: precision, recall, mean average precision, model size, and detection speed. Test results showed that the model achieved a precision of 79.93%, recall of 87.65%, mAP of 86.54%, a model size of 4.10 MB, and a detection speed of 153.46 f/s. Compared with YOLO 12s, YOLO v10s, YOLO v8s, YOLO v5s, SSD, and Faster R?CNN models, the detection speed was improved by 79.32%, 75.11%, 45.14%, 38.23%, 746.44% and 1 168.26%, respectively. Furthermore, the model maintained accurate detection performance under complex conditions such as varying lighting, partial occlusion, and image blur. When deployed on the Jetson Xavier NX embedded platform, this model achieved a mAP of 86.62%, the detection speed was 26.37 f/s. Compared with YOLO 12s and YOLO v5s, the detection speed was 90.40% and 15.51% faster, respectively. This performance was sufficient to meet the practical requirements for both accuracy and real?time processing in pollination operations. The results demonstrated that the lightweight model can accurately and efficiently identify male and female melon flowers. With only a minor loss in accuracy, it significantly reduced memory consumption, facilitating its deployment on edge devices. The research result can provide technical support for mechanical pollination of melons.