Abstract: Aiming at the problems of low efficiency, high error detection rate and poor realtime detection of apple leaf disease defects, taking gray spot, scab, rust and leaf spot disease of apple leaves as the research objects. An apple leaf disease detection algorithm based on improved yolov4 was proposed. Firstly, data amplification was used to expand the data set to improve the robustness. The algorithm determined the anchor frame by binary kmeans clustering algorithm to solve the problem that the preset anchor frame was not suitable for apple leaf diseases. DenseNet121 was introduced as a feature extraction network to improve the detection performance of apple leaf disease defects, and reduce the model size, reduce the storage overhead. The model in this paper was compared with YOLOV4 model. The results showed that the mean accuracy (mAP) of the improved YOLOv4 model reached 97.52%, which was 089% higher than before. The size of the model was 62.71 MB, which decreased 182.82 MB compared with that before improvement. The detection speed was 26.33 FPS, which increased 6.78 FPS compared with that before improvement. It can meet the demand of apple leaf disease detection in real life.

Key words: apple leaf diseases, defect detection, YOLOv4, binary K-means clustering, DenseNet121

摘要： 针对苹果叶部病害缺陷检测效率低下、误检率高、实时性差等问题，以苹果叶部的灰斑病、黑星病、锈病、斑点落叶病作为研究对象，提出一种基于改进YOLOv4算法的苹果叶部病害缺陷检测算法。首先通过数据扩增对数据集扩充提升鲁棒性，算法通过二分K均值聚类算法确定锚框以解决预设锚框不适用苹果叶部病害的问题，引入DenseNet121作为特征提取网络，提升对苹果叶部病害缺陷的检测性能，并且减小模型大小，降低存储开销。将模型与YOLOv4模型进行对比验证，试验结果表明，改进后的YOLOv4模型平均精度均值（mAP）达到97.52%，与改进前相比提升0.89%，模型大小为62.71 MB，与改进前相比减小182.82 MB，检测速度为26.33 FPS，与改进前相比提升6.78 FPS。能够满足实际生活中对苹果叶部病害检测的需求。

关键词: 苹果叶部病害, 缺陷检测, YOLOv4, 二分K均值聚类, DenseNet121