基于改进YOLO v7的苹果叶片病害检测方法

Apple Leaf Disease Detection Method Based on Improved YOLO v7

摘要:

针对苹果叶片疾病形态多样、分布密集，导致检测精度不高的问题，提出了一种改进的YOLO v7模型。首先，用双向特征金字塔网络（BiFPN）替代YOLO v7中原有的特征融合方法，以提高模型对苹果叶片上不同尺度病害的检测能力。其次，在YOLO v7的ELAN和E-ELAN模块之后，增加高效通道注意力机制（ECA），以增强模型对苹果叶片病害特征的提取能力，并提高检测精度。最后，将YOLO v7的损失函数改为SIOU损失函数，以加快模型的收敛速度。实验结果表明：改进YOLO v7模型精确率为89.4%，召回率为81.5%，mAP@0.5为90.5%，mAP@0.95为62.1%，与原始YOLO v7模型相比，分别提高4.9、5.2、3.5、4.6个百分点。改进YOLO v7模型与Faster R-CNN、SSD、YOLO v3、YOLO v5s、YOLO v7模型相比，mAP@0.5分别提升40.9、20.3、4.0、2.3、3.5个百分点，单幅图像检测时间为12ms。

关键词:苹果叶片;病害检测;YOLO v7;多尺度融合;注意力机制

Abstract:

Apples have become one of the most popular fruits in the world, and the annual production of apples in China has continued to increase. However, there are certain diseases in the growth process of apple trees, which will affect the quality and yield of apples, resulting in economic losses of fruit farmers. Therefore, in view of the problem that apple leaf diseases have diverse forms and dense distribution, resulting in low detection accuracy, an improved YOLO v7 model was proposed to accurately detect apple leaf diseases. Firstly, bidirectional feature pyramid network (BiFPN) was used to replace the original feature fusion method in YOLO v7 to improve the model’s detection ability of different scale diseases on apple leaves. Secondly, after the ELAN and E-ELAN modules of YOLO v7, an efficient channel attention mechanism (ECA) was added to enhance the ability of the model to extract features of apple leaves disease and improve detection accuracy. Finally, the loss function of YOLO v7 was changed to the SIOU loss function to accelerate the convergence speed of the model. Experimental results showed that the improved YOLO v7 model had a precision of 89.4%, a recall rate of 81.5%, a mean average precision (mAP@0.5) of 90.5%, and a mean average precision (mAP@0.95) of 62.1%. Compared with the original YOLO v7 model, they were increased by 4.9, 5.2, 3.5, and 4.6 percentage points, respectively. Compared with the Faster R-CNN, SSD, YOLO v3, YOLO v5s, and YOLO v7 models, the mAP@0.5 of improved YOLO v7 model was increased by 40.9, 20.3, 4.0, 2.3 and 3.5 percentage points, respectively, and the single image detection speed reached 12ms. The research can provide a feasible technical means for accurately detecting apple leaf diseases.

Key words:apple leaf;disease detection;YOLO v7;multi-scale fusion;attention mechanism