【摘要】 为了提高咖啡的质量和产量，实现咖啡叶病虫害高效精准检测，本文提出了一种基于改进YOLOv5s的咖啡叶病虫害识别方法。利用CARAFE上采样模块替换原始模型中的上采样模块，减少上采样过程中特征信息的损失，提升特征金字塔网络性能；在检测头前端引入GAM全局注意力机制，提取空间和通道不同维度的交互信息，增强病虫害的识别能力；用高效解耦头（Decoupled Head）替换原始耦合头区分回归和分类，加快模型收敛和提高检测精度；结果显示，改进后的模型准确率、召回率和平均精度均值分别为91.6%、86.0%、91.4%；比原始YOLOv5s平均精度均值提升了2.9%，精确度和召回率分别提升了3.1%、2.7%；与当前主流的Faster R-CNN、SSD、YOLOv4-tiny、YOLOX和YOLOv7等模型相比，平均精度均值分别提升了8.2%、20.2%、37.7%、5.9%、9.7%。本文的方法对咖啡叶病虫害检测具有较高的准确率，可以为咖啡叶病虫害检测提供参考和依据。更多还原

【Abstract】 In order to improve the quality and yield of coffee and detect coffee leaf pests and diseases efficiently and accurately, we studied the method based on improved YOLOv5s. CARAFE up-sampling module is used to replace the up-sampling module in the original model to reduce the loss of feature information in the sampling process and improve the performance of the feature pyramid network. The GAM global attention mechanism was introduced at the front end of the detection head to extract the interactive information of different dimensions of space and channel to enhance the pests and diseases identification ability. The efficient decoupling head replaced the original coupling head to distinguish regression and classification to improve model convergence and detection accuracy. The results show that the average accuracy, recall and average accuracy of the improved model are 89.6%, 86.0% and 91.4%, respectively. Compared with the average accuracy of the original YOLOv5s, the average accuracy is increased by 2.9%, and the accuracy and recall rate are increased by 3.1% and2.7%, respectively. Compared with the current mainstream models such as Faster R-CNN, SSD, YOLOv4-tiny, YOLOX and YOLOv7, the average accuracy is increased by 8.2%, 20.2%, 37.7%, 5.9% and 9.7%, respectively. Our method has a high accuracy rate for coffee leaf diseases and insect pests detection, and can provide a reference and basis for the detection of coffee leaf diseases and insect pests.更多还原