基于RetinaNet的红枣果实分类检测研究

Detection Model of Jujube Fruit Based on RetinaNet

为解决不同成熟度红枣图像中多尺度红枣目标检测问题,以自然场景下获取的红枣图像为研究对象,首先建立包含不同成熟度红枣图像数据集;然后以骨干网络ResNet50和特征金字塔网络作为特征提取器,连接2个相似结构的分类子网和回归子网,以Focal Loss为损失函数,建立基于RetinaNet的红枣成熟度分类检测模型。结果表明,基于RetinaNet红枣成熟度检测模型对红枣4种成熟度分类检测平均精度均值为74.235%,满足农业生产基本要求,该研究为智能检测红枣果实及自动化采摘可行性提供了技术参考。

In order to solve multi⁃scale object detection problem of jujube fruits with varied maturity under natural environment,a detection network was proposed to rapidly detect the jujube fruits.Firstly,jujube image dataset taken in natural environment was established.Then,the model was constructed with backbone network ResNet50 and Feature Pyramid network(FPN)as feature extractors,and connected with classi⁃fication subnets and regression subnets,and Focal Loss was used as loss function.Finally a detection model for four maturity categories of jujube fruits based on RetinaNet was established.The mean Average Precision of RetinaNet model was 74.235%.The results showed that the detection model for jujube fruits maturity based on RetinaNet met the basic requirements of agricultural production.This study provided a technical refer⁃ence for the feasibility of intelligent detection and automatic picking of jujube fruits.