Abstract: Rice leaf disease prevention plays an important role in improving rice yield. Aiming at the problems of slow manual inspection speed and high subjectivity of rice leaf disease, a target detection method of rice leaf disease based on improved Yolov5s is proposed. The Kmeans clustering algorithm is used to obtain the prior frame size, which enhances the adaptability of the detection model to rice leaf disease. The lightweight spatial attention and channel attention are fused to enhance the highlevel semantic feature information and the models awareness of disease information. Finally, the feature pyramid network is combined with the multiscale receptive field to obtain target context information, which effectively enhances the models extraction of features around the target and improves the accuracy of target detection. The experimental results show that the average detection accuracy (IOU=0.5) of the improved Yolov5s algorithm is increased by 4.3%, the F1 value is increased by 5.3%, and the FPS is 58.7 f/s. The proposed method effectively improves the detection accuracy of the Yolov5s algorithm for rice leaf disease and meets the demand of realtime detection.

Key words: rice leaf disease detection, Kmeans clustering, attention mechanism, multiscale receptive field

摘要： 水稻叶病防治在提高水稻产量中具有重要作用，针对水稻叶病人工检查速度慢、主观性高的问题，提出一种基于改进Yolov5s的水稻叶病目标检测方法。采用Kmeans聚类算法得到先验框尺寸，增强检测模型对水稻叶病的适应性；将轻量级空间注意力与通道注意力融合，对高层语义特征信息增强，增强模型对病害信息的感知度；并结合特征金字塔网络，融合多尺度感受野获取目标上下文信息，有效地增强模型对目标周围特征的提取，提高目标检测的准确度。试验结果表明：改进后的Yolov5s算法平均检测精度(IOU=0.5)提高4.3%，F1值提高5.3%，帧率FPS为58.7 f/s。有效提升Yolov5s算法对水稻叶病的检测精度，达到实时检测的需求。

关键词: 水稻叶病检测, Kmeans聚类, 注意力机制, 多尺度感受野