基于MHSA-YOLOv7的小麦赤霉病感染率检测

Detection of Gibberella Infection Rate in Wheat Based on MHSA-YOLOv7

在抗病育种中小麦赤霉病感染率是衡量籽粒抗性表型鉴定的重要衡量指标,针对目前小麦赤霉病感染率检测存在检测时间长、硬件成本高以及检测方式破坏植株等问题,设计了一种适用于麦穗籽粒此类小目标检测的深度学习网络模型——MHSA-YOLOv7。通过在原YOLOv7主干网络中融合多头自注意力(Muti-Head Self-Attention,MHSA)机制来提高模型对深层语义特征的提取能力,并使用加权双向特征金字塔网络(Bidirectional Feature Pyramid Network,BiFPN)实现模块间的跨层连接,使该模型能够提取和传递更丰富的特征信息。实验结果表明,MHSA-YOLOv7在小麦单穗赤霉病数据集上达到了90.75%的检测精度,相较于原YOLOv7模型,改进后的算法对于麦穗籽粒此类小目标物体具有更强的特征提取能力,检测精度、召回率、F1值、mAP@0.5以及mAP@0.5∶0.95分别提高了0.33%、1.83%、0.011、1.19%以及0.38%,有效满足了小麦赤霉病感染率的精确检测,为实现小麦植株病害走势的长期观测以及小麦籽粒抗性的准确评估提供了技术支持。

In disease resistance breeding,the infection rate of gibberella in wheat is an important indicator to measure the phenotype identification of grain resistance.In view of the problems of long detection time,high hardware cost and damage to plants in the detection of wheat gibberella infection,a deep learning network model,or MHSA-YOLOv7 suitable for the detection of small objects such as wheat ear grain is designed.By integrating the Muti-Head Self-Attention(MHSA)mechanism in the original YOLOv7 backbone network,the model can extract deep semantic features,and the weighted Bidirectional Feature Pyramid Network(BiFPN)is used to realize the cross-layer connection between modules,so that the model can extract and transmit richer feature information.The experimental results show that MHSA-YOLOv7 achieves a detection accuracy of 90.75%on the wheat single ear gibberella dataset.Compared with the original YOLOv7 model,the improved algorithm has stronger feature extraction ability for small objects such as wheat ear grain,and the detection Accuracy,Recall,F1 score,mAP@0.5 and mAP@0.5:0.95 are improved by 0.33%,1.83%,0.011,1.19%and 0.38%respectively.The improved algorithm effectively satisfies the accurate detection of wheat gibberella infection rate,and provides technical support for long-term observation of wheat disease trends and accurate assessment of wheat grain resistance.