便携式柑橘虫害实时检测系统的研制与试验

为实现柑橘虫害的快速、准确识别,帮助果农及时掌握果园内虫害的危害程度和分布情况,该研究结合嵌入式图像处理技术设计了一套基于深度卷积神经网络的柑橘虫害实时检测系统。优选MoblieNet作为虫害图像特征提取网络,区域候选网络生成害虫的初步位置候选框,快速区域卷积神经网络(Faster Region Convolutional Neural Networks,Faster R-CNN)实现候选框的分类和定位。检测系统根据目标图像中虫害数量计算危害程度,按照正常、轻度、中度、重度4个等级判定柑橘虫害的严重程度,形成虫害识别与级别定量化测评软件。最后引入北斗模块获取采样点位置信息,进一步处理成可视化的虫害热力图。结果表明,该方法可实现对柑橘红蜘蛛和蚜虫的快速准确检测,识别准确率分别达到91.0%和89.0%,单帧图像平均处理速度低至286ms。该系统实现了柑橘虫害的精准识别与定位,可为农药喷洒作业提供精准信息服务。

Development and experiment of the portable real-time detection system for citrus pests

In order to achieve a rapid, accurate and non-destructive detection of citrus pest infestation levels of the fruit trees, a real-time citrus pest detection system based on deep convolutional neural network was designed and developed in this study. The system was composed of a perception layer, a network layer and an application layer. The perception layer was responsible for the collection and identification of pest image data; the network layer was responsible for the data encoding, authentication and transmission between the detection instrument and the cloud server, and between the cloud server and the client; the application layer calculated the degree of damage resulted from pests, based on the number of pests in the target image, then the Beidou module was introduced to obtain the location information of the sampling points, and finally a visual pest heat map was generated. In order to obtain a pest recognition model suitable for the computing requirements of embedded devices, MoblieNet was preferred as the pest image feature extraction network. The regional candidate network generated the preliminary position candidate frame of the pests, and Faster Region Convolutional Neural Networks (Faster R-CNN) realized the classification and positioning of the candidate frame. The results showed that, compared with VGG16 and GoogleNet feature extraction network, MobileNet had the smallest parameter amount, only 15.147 M, and the mean average precision (mAP) and accuracy (ACC) indicators were 86.40% and 91.07%, respectively. Although the mAP and ACC of MobileNet were lower than VGG16 and higher than GoogleNet, the average time of MobileNet to process an image was 286 ms, which was much less than VGG16 (679 ms) and GoogleNet (459 ms). Considering comprehensively, MobileNet was used as the feature extraction network of the citrus pest detection model in this study. According to the detection effect of the two citrus pests, spider mite and aphids, the recognition rates were both high, reaching 91.0% and 89.0%, respectively, indicating that the sample features were selected correctly. From the perspective of counting accuracy, spider mite was 90.1%, and that of aphids was 43.8%. The main reason was that aphids were dense and obscure each other, and there was overlap so that it was difficult to label all pests when labeling. During the model training process, some aphids samples became negative samples and the accuracy rate was reduced. In addition, the pest distribution heat map had a small error, and directly displayed the degree of damage in different target point. The system realized the accurate identification and positioning of citrus pests, and provided an accurate information services for pesticide spraying operations.