Abstract: Aiming at the problems of low accuracy，slow model running speed and no detection function of the existing tea disease identification methods，a novel Yolov7‐TSA（Yolov7‐Tiny‐SiLU‐Attention）lightweight network architecture was proposed to detect and classify tea diseases.Specifically，the LeakReLU activation function in the Yolov7‐T network was replaced with the SiLU activation function to improve detection accuracy and prevent overfitting.Meanwhile，the feature perception of target contour and spatial location was further improved by fusing the coordinate attention mechanism. Experiments on the dataset containing eight tea diseases（including healthy tea leaves） showed that the recognition accuracy of Yolov7‐TSA network reached 94.2%，which was 3.2 and 1.2 percentage points higher compared to Yolov7‐T and Yolov7 networks，respectively.Furthermore，it showed significant results in terms of parameters，floating point operations，model size and inference time per image，which were reduced by 83%，87%，83% and 34%，respectively，compared to the Yolov7 network. The network model achieves the detection and classification of tea diseases when balancing recognition accuracy and real‐time performance.

Key words: Tea disease, Disease detection, Disease classification, Yolov7‐TSA, Coordinate attention, Lightweight network, Computer vision, Smart agriculture

摘要： 针对现有茶叶病害识别方法准确率低、模型运行速度慢和缺乏检测功能等问题，提出一种新的Yolov7-TSA（Yolov7-Tiny-SiLU-Attention）轻量型网络架构对茶叶病害检测和分类。将Yolov7-T网络中的LeakReLU激活函数替换为SiLU激活函数，以提升检测精度，并防止过拟合。同时，通过融合坐标注意力机制进一步提升对目标轮廓和空间位置的特征感知能力。在含8种茶叶病害（含健康茶叶）的数据集上试验。结果表明，Yolov7-TSA网络对茶叶病害的识别准确率达到了94.2%，与Yolov7-T、Yolov7网络相比，分别提升了3.2、1.2个百分点。另外，Yolov7-TSA网络在参数量、浮点运算数、模型大小和单张图片推理时间方面表现出了显著的效果，与Yolov7网络相比，其分别降低了83%、87%、83%和34%。该网络模型实现了对茶叶病害的检测与分类，同时平衡了识别准确率和实时性能。

关键词: 茶叶病害, 病害检测, 病害分类, Yolov7-TSA, 坐标注意力, 轻量型网络, 计算机视觉, 智慧农业