摘要
为了解决现有的农作物病害检测方法对不同番茄叶片病害检测的精度低、效果差的问题,提出一种基于YOLOv5网络模型改进的番茄叶片病害检测模型YOLOv5s-TLD。首先在原YOLOv5s模型的Backbone中构建DCAM注意力机制模块,通过制定双通道注意力和空间注意力机制加强模型对番茄叶片病理特征的提取能力,并减弱模型受复杂背景特征的影响,以提高模型对不同种类病害的检测精度和分类精度;然后应用融合Swin Transformer的C3STR模块替换原网络第6层的C3模块,强化模型在多尺度上建模的能力,实现模型对小尺寸的番茄叶片病害残差特征的高精度学习;再运用BiFPN加权双向特征金字塔网络替换原YOLOv5模型Head的PANet路径聚合网络,该网络采用跨尺度特征融合和可学习权重的方式融合模型不同层次的特征,在增强网络的特征融合能力的同时使网络获得更多的特征信息,以提高模型的感受野和特征表达能力;最后进行不同模型的检测对比试验,并在实际复杂场景下进行番茄叶片病害检测试验。试验结果表明:YOLOv5s-TLD模型平均精度均值和召回率分别为97.7%和96.3%,较原YOLOv5s模型平均精度均值和召回率分别提高1.9个百分点和2.5个百分点。该模型具有良好的检测精度和检测效果,且该模型在背景复杂的实际种植环境下能够准确地检测并识别不同种类的番茄叶片病害,研究结果可为农业智能管理和番茄叶片病害检测技术的实际应用提供参考。
In order to solve the problem of low accuracy and poor effect of existing crop disease detection methods for different tomato leaf diseases,a tomato leaf disease detection model YOLOv5s-TLD based on YOLOv5 network model was proposed.Firstly,the DCAM attention mechanism module was constructed in the Backbone of the original YOLOv5s model.Dual-channel attention and spatial attention mechanisms were developed to strengthen the model's ability to extract pathological features of tomato leaves and weaken the model's influence on complex background features,to improve the model's detection accuracy and classification accuracy for different kinds of diseases.Secondly,the C3STR module of integrated Swin Transformer was used to replace the C3 module at the sixth layer of the original network to strengthen the model's multi-scale modeling ability and realize the model's high-precision learning of small-size tomato leaf disease residual features.Then the BiFPN weighted bidirectional feature pyramid network was used to replace the PANet path aggregation network of the Head of the original YOLOv5 model.The network used cross-scale feature fusion and learnable weights to integrate features of different levels of the model,which enhanced the feature fusion capability of the network and enabled the network to obtain more feature information to improve the model's receptive field and feature expression ability.Finally,different models were tested and compared,and tomato leaf disease test was carried out in the actual complex scene.The experimental results showed that the average accuracy and recall rate of the YOLOv5s-TLD model were 97.7%and 96.3%,respectively,which were 1.9 percentage points and 2.5 percentage points higher than the original YOLOv5s model.The model has good detection accuracy and detection effect,and the model can accurately detect and identify different types of tomato leaf diseases under the actual growing environment with complex background.The research results can provide references for the practical applica
作者
陶兆胜
石鑫宇
王勇
伍毅
吴浩
TAO Zhao-sheng;SHI Xin-yu;WANG Yong;WU Yi;WU Hao(College of Mechanical Engineering,Anhui University of Technology,Maanshan Anhui 243032,China)
出处
《沈阳农业大学学报》
CAS
CSCD
北大核心
2023年第6期712-721,共10页
Journal of Shenyang Agricultural University
基金
安徽省自然科学基金面上项目(2108085ME166)
安徽高校自然科学研究项目重点项目(KJ2021A0408)。
