基于改进ResNet34模型的杂草识别方法

Weed recognition method based on modified ResNet34 model

要实现杂草的精准防控,除草剂的变量喷洒,杂草的准确识别是前提。针对田间复杂的自然环境,为了进一步提高杂草识别准确率,解决模型在识别中泛化和拟合能力较差等问题,提出了一种基于改进ResNet34模型的杂草识别方法。该方法以ResNet34为基础模型,将高效通道注意力机制(Efficient Channel Attention,ECA)与残差块相结合作为骨干网络,以此优化网络权重,强化对目标杂草的特征提取。并采用空洞卷积增加引入前方特征图的感受野,减少特征图精度上的损失。实验结果表明,该方法较未改进的AlexNet,GoogLeNet,VGG16,ResNet34网络相比,识别准确率分别提高了12%,16.41%,10.62%,1.44%,同时泛化能力更强。该方法也为杂草防治提供了新的解决方案,为农田的智能除草提供了新思路。

To achieve accurate weed control,variable spraying of herbicides and accurate identification of weeds are prerequisites.For the complex natural environment in the field,a weed recognition method based on the improved ResNet34 model is proposed to further improve the weed recognition accuracy and to solve the problems of poor generalization and fitting ability of the model in recognition.The method takes ResNet34 as the base model and combines Efficient Channel Attention(ECA)mechanism with residual blocks as the backbone network,as a way to optimize the network weights and enhance the feature extraction of the target weeds.And the dilated convolution is used to increase the perceptual field of the feature map introduced in front to reduce the loss in feature map accuracy.The experimental results show that this method improves the recognition accuracy by 12%,16.41%,10.62%,and 1.44%compared with the unimproved AlexNet,GoogLeNet,VGG16,and ResNet34 networks,respectively,and also has better generalization ability.The method also provides a new solution for weed control and a new idea for intelligent weed control in agricultural fields.