基于改进YOLOv3网络模型的茶草位置检测算法

精准高效的茶草识别是智能茶园植保机械进行除草工作的关键。针对目前茶园除草智能化程度较低等问题，提出改进YOLOv3网络模型的茶草检测算法。首先，分季节和时间段，在多个茶叶品种的种植园中以自适应的距离和角度采集茶草混合图像并建立试验数据集。接着，使用K均值聚类算法重新设计先验锚框尺度。然后，以YOLOv3网络模型为基础，选取17×17的网格划分图像区域；采用残差网络（ResNet）作为主干网；加入过程提取层，增强草株检测性能。最后在原损失函数中引入广义交并比损失。通过消融试验和不同目标检测算法对比试验验证此改进算法对茶树与杂草的检测效果。试验结果表明，改进 YOLOv3网络模型对杂草的检测精确率和召回率分别为85.34%和91.38%，对茶树的检测精确率和召回率最高达到82.56%和90.12%；与原YOLOv3网络模型相比，检测精确率提高8.05%，并且每秒传输帧数达到52.83 Hz，是Faster R-CNN网络模型的16倍。这些数据说明所提算法在茶园复杂环境下，不仅对于茶树和杂草具有更好的识别效果，而且满足实时检测的要求，可以为智能茶园植保机械提供技术支持。

An algorithm for localizing tea bushes and green weeds based on improved YOLOv3 network model

Precise and efficient tea weed identification is the key to weed control of intelligent tea plantation protection machinery. In response to the current problems of low level of weeding intelligence in tea gardens, a tea weed detection algorithm based on the improved YOLOv3 network model is proposed. Firstly, during different seasons and periods, tea weed images are collected with a self adaptive distance and angle in the plantations of multiple tea varieties and to build experimental data sets. Secondly, the prior anchor box scales are redesigned by the K means clustering algorithm. Then, based on the YOLOv3 network model, an image area is divided by the grids of 17×17, the residual network (ResNet) is utilized as the backbone network, and the process extraction layer is added to it to improve the detection performance for weeds. Finally, the generalized intersection over union loss is introduced in the original loss function. The effectiveness of the improved algorithm is verified for tea weed detection via ablation study and comparison experiment of different target detection algorithms. The experimental results show that the detection precision and recall rate of the improved YOLOv3 network model for weeds are 85.34% and 91.38%, respectively, the highest detection precision and recall rate of tea bushes are 82.56% and 90.12%, respectively; compared with the original YOLOv3 network model, the precision is improved by 8.05%, and the frames per second transmission reaches 52.83 Hz, 16 times of the Faster R-CNN network model. These datas demonstrate that proposed algorithm in the complex environment of tea plantations not only provides better detection effect for tea bushes and green weeds, but also satisfies the requirement of real time detection, which can provide technical support for intelligent tea plantation machinery.