基于PP-YOLO深度学习模型的赣南脐橙目标检测方法

果实检测在研究脐橙采摘机械化发展中有着重要作用，然而不良天气条件将对目标果实的检测和识别产生不利影响。针对雾天和雨天情形下脐橙果实图像模糊、噪声复杂，检测速度较慢和准确率较低的问题，通过采用单阶段目标检测网络PP-YOLO来研究不良天气条件下赣南脐橙果实的识别。通过主干网络ResNet提取特征并结合FPN（特征金字塔网络）进行特征融合实现多尺度检测，且基本实现端到端检测。实验结果表明，所提出的PP-YOLO检测模型可实现雾天和雨天情况下赣南脐橙检测任务，mAP分别为89.06%和91.01%，识别效率分别可达到75.30 fps和75.44 fps，可以尝试在脐橙采摘机器人的研制中加以应用。

FRUIT RECOGNITION METHOD OF GANNAN NAVEL ORANGE BASED ON PP-YOLO DEEP LEARNING MODEL

Fruit detection plays an important role in studying the development of navel orange picking mechanization. However, adverse weather conditions will have an adverse impact on the detection and identification of target fruits. Aiming at the problems of blurry images, complex noise, slow detection speed and low accuracy rate of navel orange fruit under foggy and rainy days, this paper uses a single-stage target detection network PP-YOLO to study the identification of Gannan navel orange fruit under bad weather conditions. Feature extraction is achieved by the backbone network ResNet and feature fusion by combining FPN (feature pyramid network), and end-to-end detection is basically realized. The experimental results show that the proposed PP-YOLO detection model can realize the Gannan navel orange detection task under fog and rainy days, the mAP is 89.06% and 91.01%, and the recognition efficiency can reach 75.30 and 75.44fps, respectively, which can be tried to be applied in the development of navel orange picking robot.