Abstract: In natural scenes, tea diseases have different shapes and small targets, and traditional convolutional neural networks are not suitable for disease detection under complex background. Therefore, we proposed an improved tea disease target detection algorithm, YOLOv5sECAASFF. This algorithm introduces ECA channel attention module to enhance the global context information in the channel dimension, and uses adaptive spatial feature fusion (ASFF) technology to improve the multiscale feature fusion of tea diseases and improve the background antijamming ability of the model. At the same time, the GIoU loss function is used as the bounding box loss function to further improve the detection accuracy of the regression target. Compared with the original YOLOv5s model, the average accuracy of the improved YOlOv5SECAASFF model in the identification of tea white star disease, tea wheel spot disease, tea anthracnose disease and tea algal spot disease was increased by 5%, 4%, 3% and 2%, respectively, and the average accuracy was 92.1%. In addition, the image detection speed of this model is 64 f/s, and the comprehensive performance is better than that of YOLOv4, SSD and Faster RCNN models. Therefore, the model provides a reference for the detection of different kinds of tea diseases in the natural growing environment, and provides important technical support for early prediction.

Key words: deep learning, YOLOv5s, tea diseases, attention mechanism, target detection

摘要： 在自然场景下，茶叶病害形状各异、目标小，传统卷积神经网络不适用于复杂背景下的病害检测。因此，提出一种改进的YOLOv5sECAASFF茶叶病害目标检测算法。该算法通过引入ECA通道注意力模块来增强通道维度上的全局上下文信息，并使用自适应空间特征融合(ASFF)技术改进茶叶病害的多尺度特征融合，提高模型的背景抗干扰能力。同时，使用GIoU损失函数作为边界框损失函数，进一步提高回归目标的检测精度。与原始YOLOv5s模型相比，改进YOLOv5sECAASFF模型在茶白星病、茶轮斑病、茶炭疽病、茶藻斑病的识别平均精度分别提高5%、4%、3%、2%，平均精度均值为92.1%。此外，该模型的图像检测速度为64f/s，并且综合性能也优于YOLOv4、SSD和Faster RCNN模型。因此，该模型为茶叶在自然生长环境下不同种类病害的检测提供参考，并为早期预测提供重要的技术支持。

关键词: 深度学习, YOLOv5s, 茶叶病害, 注意力机制, 目标检测