TTLD-YOLOv7:非结构化环境下茶树病害的检测算法  

TTLD-YOLOv7:An Algorithm for Detecting Tea Diseases in An Unstructured Environment

作　　者：俞淑燕 杜晓晨[1] 冯海林[1] 李颜娥 YU Shuyan;DU Xiaochen;FENG Hailin;LI Yan′e(College of Mathematics and Computer Science,Zhejiang Agriculture and Forestry University,Hangzhou 311300,China)

机构地区：[1]浙江农林大学数学与计算机科学学院,浙江杭州311300

出　　处：《茶叶科学》2024年第3期453-468,共16页Journal of Tea Science

基　　金：浙江省“三农九方”科技协作项目(2022SNJF036)。

摘　　要：茶树病害对茶树种植业和相关行业的影响极为严重。在动态而复杂的茶园环境中检测疾病的传统方法效率低下,检测效果不尽人意。本研究提出一种基于YOLOv7-tiny的模型,增强了茶树病害的细微检测能力。通过整合CoordConv和ECA信道关注机制,本模型在卷积特征图中实现了更高的空间识别能力,并降低了背景噪声对特征识别的影响。进一步的改进包括采用归一化瓦瑟斯坦距离度量和去耦头,以提高对小病斑的检测能力。使用K-means算法根据茶树病斑的特殊性生成了新的锚框,提高了模型的精确性和通用性。对比分析表明,该模型优于现有模型FasterR-CNN、 SSD、 YOLOv5s、 YOLO-Tea、 YOLOv7-tiny和YOLOv7,平均精确度提高5.39个百分点,达到了93%。改进后的模型可应用于茶树病害监测。Tea diseases have an extremely serious impact on tea plantations and related industries.Traditional methods for disease detection in the dynamic and complex tea plantation environment are inefficient and unsatisfactory.This study proposed that a YOLOv7-tiny-based model enhanced the fine-grained detection of tea tree diseases.By integrating CoordConv and ECA channel attention mechanisms,this model achieved higher spatial recognition capability in convolutional feature maps and reduced the effect of background noise on feature recognition.Further improvements included the use of a normalized Wasserstein distance metric and decoupled heads to improve the detection of small spots.A new anchor frame was generated using the K-means algorithm based on the specificity of tea spots to improve the accuracy and generalizability of the model.Comparative analysis shows that the model outperforms the existing models Faster R-CNN,SSD,YOLOv5s,YOLO-Tea,YOLOv7-tiny,and YOLOv7,with an average accuracy improvement of 5.9 percentage points to 93%.The improved model could be applied to tea disease monitoring.

关 键 词：茶树病害 YOLOv7-tiny 自然环境 目标检测 

分 类 号：S571.1[农业科学—茶叶生产加工]