摘要
为解决传统识别方法效率低、准确率不佳、相关研究不足等问题,提出一种基于改进Resnet34的小麦开花期品种识别模型。首先,针对现有农业识别模型参数量较多,不利于在移动端部署的问题,使用改进Inceptionv1模块替代Resnet34网络基本残差块的第2个卷积块,使模型参数量降低了一半左右;其次,针对模型参数量减少后识别准确率下降的问题,在模型中加入ECA与simAM注意力机制,以期通过对小麦特征的有效提取提升小麦开花期品种识别准确率。实验结果表明,所提模型在小麦开花期数据集上的平均识别准确率达95.7%,相较原始Resnet34模型提高了2.1%,相较efficientnetv2_s、MobileNet-v2、GoogLeNet模型准确率分别提高了2.4%、3.2%、5.0%。所提模型具有更好的特征提取能力,为小麦开花期品种识别提供了一种有效方法。
To solve the problems of low efficiency,low accuracy,and insufficient related research in traditional recognition methods,a wheat flowering period variety recognition model based on improved Resnet34 is proposed.Firstly,to address the problem that existing agricultural recognition models have a large number of parameters that are not conducive to deployment on mobile devices,an improved Inceptionv1 module is used to replace the second convolutional block of the basic residual block of the Resnet34 network,reducing the model parameter count by about half;Secondly,in response to the problem of decreased recognition accuracy after the reduction of model parameters,ECA and simAM attention mechanisms are added to the model to improve the accuracy of wheat flowering stage variety recognition through effective extraction of wheat features.The experimental results show that the proposed model has an average recognition accuracy of 95.7%on the wheat flowering stage dataset,which is 2.1%higher than the original Resnet34 model.Compared with the efficientnetv2_s,MobileNet-v2,and GoogLeNet models,the accuracy has been improved by 2.4%,3.2%,and 5.0%,respectively.The proposed model has better feature extraction ability and provides an effective method for identifying wheat varieties during the flowering period.
作者
冯永强
刘成忠
韩俊英
邢雪
杨红强
FENG Yongqiang;LIU Chengzhong;HAN Junying;XING Xue;YANG Hongqiang(College of Information Sciences and Technology,Gansu Agricultural University,Lanzhou 730070,China)
出处
《软件导刊》
2024年第5期1-8,共8页
Software Guide
基金
国家自然科学基金项目(32360437)
甘肃省高等学校创新基金项目(2021A-056)
甘肃省高等学校产业支撑计划项目(2021CYZC-57)。
