摘要
为解决自然环境中番茄叶片病虫害检测场景复杂、检测精度较低,计算复杂度高等问题,该研究提出一种SLPYOLOv7-tiny的深度学习算法。首先,将主干特征提取网络中部分3×3的卷积Conv2D(2D convolution)改为分布偏移卷积DSConv2D(2D Depthwise Separable Convolution),以减少网络的计算量,并且使计算速度更快,占用内存更少;其次,将无参数注意力机制(parameter-free attention module,SimAM)融合到骨干特征提取网络中,加强模型对病虫害特征的有效提取能力和特征整合能力;最后,将原始YOLOv7-tiny的CIOU损失函数,更替为Focal-EIOU损失函数,加快模型收敛并降低损失值。试验结果表明,SLP-YOLOv7-tiny模型整体识别精准度、召回率、平均精度均值mAP_(0.5)(IOU阈值为0.5时的平均精度)、mAP_(0.5~0.95)(IOU阈值从0.5到0.95之间的平均精度)分别为95.9%、94.6%、98.0%、91.4%,与改进前YOLOv7-tiny相比,分别提升14.7、29.2、20.2、30个百分点,同时,计算量降低了62.6%。与YOLOv5n、YOLOv5s、YOLOv5m、YOLOv7、YOLOv7-tiny、Faster-RCNN、SSD目标检测模型相比,mAP_(0.5)分别提升了2.0、1.6、2.0、2.2、20.2、6.1和5.3个百分点,而计算量大小仅为YOLOv5s、YOLOv5m、YOLOv7、Faster-RCNN、SSD的31.5%、10.6%、4.9%、4.3%、3.8%。结果表明SLP-YOLOv7-tiny可以准确快速地实现番茄叶片病虫害的检测,且模型较小,可为番茄叶片病虫害的快速精准检测的发展提供一定的技术支持。
In order to solve the problems of complex detection scene,low detection accuracy and high computational complexity of tomato leaf pests and diseases in natural environment,a deep learning algorithm SLP-YOLOv7-tiny was proposed.Firstly,part of the 3×3 convolution Conv2D Convolution in the backbone feature extraction network is changed to distributed migration convolution DSConv2D(2D Depthwise Separable Convolution)to reduce the computing load of the network and speed it up.Less memory usage;Secondly,the parameter-free attention module(SimAM)was integrated into the backbone feature extraction network to enhance the model's ability to effectively extract and integrate features of pests and diseases.Finally,the original YOLOv7-tiny CIOU loss function is replaced with Focal EIOU loss function to accelerate the model convergence and reduce the loss value.The test results show that The overall identification accuracy,recall rate,average accuracy mAP_(0.5)(the average accuracy when the IOU threshold is 0.5),and mAP_(0.5~0.95)(the average accuracy when the IOU threshold between 0.5 and 0.95)of SLP-YOLOv7-tiny model are 95.9%,94.6%,and 98%,respectively Compared with YOLOv7-tiny before the improvement,they respectively increase by 14.7,29.2,20.2 and 30 percentage points.Meanwhile,the calculation amount decreases by 62.6% and the detection speed increases by 13.2%.Compared with YOLOv5n,YOLOv5s,YOLOv5m,YOLOv7,Yolov7-Tiny,Faster-RCNN and SSD target detection models,mAP_(0.5) improved by 2.0,1.6,2.0,2.2,20.2,6.1 and 5.3 percentage points respectively.The computing capacity was only 31.5%,10.6%,4.9%,4.3% and 3.8% of YOLOv5s,YOLOv5m,YOLOv7,father-RCNN and SSD.The results show that SLP-YOLOv7-tiny can accurately and quickly detect tomato leaf diseases and pests,and the model is small,which can provide certain technical support for the development of rapid and accurate detection of tomato leaf diseases and pests.SLP-YOLOv7-tiny can be expected to accurately and rapidly detect the tomato leaf diseases and insect pests.The small model was more conducive to the migration application.A comparison was also performed on the detection performance of different attention mechanisms.The SimAM attention mechanism shared the better screening of effective feature information on the YOLOv7-tiny model.The disease spots were identified to effectively improve the accuracy of the model,compared with the SENet and CBAM attention mechanism.The ablation test was carried out to verified the added module.The high detection accuracy and computational efficiency were suitable for the deployment in the natural environment with the limited computing resources,such as mobile terminals.The visualization of model detection show that the SLP-YOLOv7-tiny model can be used to learn and distinguish the fine features of tomato leaf disease spots.The detection accuracy was also better than that of the YOLOv7-tiny model.Moreover,the SLPYOLOv7-tiny model can be used to accurately identify the diseases and pests of diseased tomato leaves in the natural shooting environment.The high detection accuracy can be applied for the multiple diseases and pests.The better identification was also obtained for the small features in the similar diseases and pests.The finding can provide the technical support for the rapid and accurate detection of tomato leaf diseases and insect pests.
作者
王会征
孙良晨
李新龙
刘海藤
王国宾
兰玉彬
WANG Huizheng;SUN Liangchen;LI Xinlong;LIU Haiteng;WANG Guobin;LAN Yubin(College of Agricultural Engineering and Food Science,Shandong University of Technology,Zibo 255049,China)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2024年第10期194-202,共9页
Transactions of the Chinese Society of Agricultural Engineering
基金
山东省自然科学基金项目(ZR2023MC133
ZR2021QC154)
宁夏回族自治区重点研发计划项目(2023BCF01051)。
