基于机器视觉的小贯小绿叶蝉智能识别的研究与应用

深度学习已经在农作物害虫实时监测的智能识别过程中广泛应用。以小贯小绿叶蝉（Empoasca onukii）识别模型为基础,研究深度学习在诱虫板上叶蝉识别中的应用,旨在提高小贯小绿叶蝉田间种群调查的准确性。本研究设计了一种茶园小贯小绿叶蝉的识别、计数方法,首先采用黄色诱虫板诱集小贯小绿叶蝉,利用相机对诱虫板进行图像采集并上传至服务器,然后通过服务器部署的目标检测算法,对图像中叶蝉进行识别与计数。通过算法筛选,确定YOLOv3作为识别算法,用改进后的Soft-NMS代替原来的NMS,用K-means聚类方法计算新的先验框的尺寸,提升YOLOv3对目标识别的速度和准确率。通过田间试验对比诱虫板上叶蝉的真实数量,结果显示优化后识别算法的准确率可达到95.35%以上。本研究验证了诱虫板诱集、目标识别算法和物联网技术相结合,能够为小贯小绿叶蝉田间种群的实时监测提供技术支持,可为其他具有颜色偏爱性昆虫的实时监测和茶园害虫综合治理提供参考。

Research and Application of Intelligent Identification of Empoasca onukii Based on Machine Vision

Deep learning has been widely used in intelligent identification in the progress of real-time monitoring of crop pests. Based on the identification model of tea leafhopper, Empoasca onukii, the application of deep learning in field leafhopper recognition was introduced to improve the precision of field population investigation of E. onukii. In this paper, a method of identification and count of E. onukii in tea garden was designed. Firstly, yellow sticky card was used to attract tea leafhoppers, and images of cards were collected by camera and uploaded to the web server. Then, target detection algorithm deployed by the server was used to identify and count the leafhoppers in the images. Through algorithm screening, YOLOv3 was determined as the recognition algorithm, and the improved soft-NMS was used to replace the original NMS. K-means clustering method was used to calculate the size of the new prior frame, so as to improve the speed and precision of YOLOv3. The results show that the average precision of the optimized algorithm could reach more than 95.35% comparing with the real number of leafhoppers on the sticky card. Therefore, the combination of the sticky card trapping, target recognition algorithm and internet of things technology could realize the real-time monitoring of population for E. onukii, which could provide a reference for other insects with color preference and integrated pest management in tea gardens.