摘要
精准识别作物害虫是控制虫害发生态势的重要基础。针对现有害虫识别准确率较低、基于卷积神经网络的害虫识别结构较复杂且计算成本较高、害虫识别模型泛化能力低及难以部署等问题,该研究提出了一种基于改进EfficientNet模型的作物害虫智能识别模型。该模型通过引入坐标注意力(Coordinate Attention,CA)机制而改进EfficientNet主体结构,引入数据增强的组合训练策略及Adam优化算法来提高模型的泛化能力,并采用迁移学习策略来训练改进的EfficientNet模型,从而提出了一个高性能轻量化的作物害虫识别模型CA-EfficientNet。在公开的大规模作物害虫数据集IP102上展开试验,结果表明该研究提出的CA-EfficientNet模型识别准确率达到69.45%,较改进前提高了4.01个百分点;与现有同类最优算法(GAEnsemble)的性能相比,识别准确率高出2.32个百分点。改进后的CA-EfficientNet模型参数量为5.38 M,较改进前仅增加了0.09 M;相比于经典分类网络VGG、ResNet-50、GoogleNet等,其参数量仅是这些网络模型参数量的3.89%、22.72%和52.63%。试验结果表明,所提方法有效提高了作物害虫图像的识别准确率,较大幅度地减少了模型参数量,在保持轻量化计算的基础上获得了明显优于同类最优算法的准确率。
An accurate recognition of crop pests has been one of the most important steps to control the pest occurrence for the higher crop yield. It is still a great challenge to effectively determine the characteristics of crop pests, where the appearance of crop pests belonging to the same species significantly varies with the growth periods, while the morphological features of crop pests also vary in the different species resemble each other. However, the manual identification and traditional Support Vector Machine(SVM) machine learning cannot fully meet the production needs of pest recognition in modern agriculture at present.Deep learning can be widely expected to identify pest species in recent years. Nevertheless, there is a large computational cost in the current Convolutional Neural Networks(CNN) for the feature extraction, due to the complex structure, thus leading to the lower recognition accuracy on a large number of dataset. This study aims to propose a pest intelligent recognition with high-performance, lightweight, and easy to apply for the production needs of smart agriculture. An improved Efficient Net-based scheme was established for crop pest recognition. First, the Coordinate Attention(CA) mechanism was introduced into the EfficientNet network structure, further locating the Region of Interest(ROI) area in a pest image using feature location information, which in turn improved the feature representation capability of the model. Second, the combined training strategy of data augmentation was developed to improve the diversity of pest samples, the robustness, and the generalization of the model. Third, an Adam optimization was used to further improve the convergence performance of the model. Last, a transfer learning strategy was also involved to initialize the parameters of the model. As such, a deep learning network named CA-EfficientNet was established to integrate these approaches, where the public large-scale dataset IP102 was taken as the network model training and performance testing in an experimental simulation. The results show that the CA-EfficientNet reached an accuracy of 69.45%, which was 4.01 percentage points higher than before, and 2.32 percentage points larger than the state-of-the-art GAEnsemble method for pest recognition. The amount of parameters dataset was 5.38 M in the improved CA-EfficientNet, and only 3.89%, 22.72%, and 52.63% of that for the VGG, Res Net-50, GoogleNet. In summary, the scheme remarkably improved the accuracy of recognition for a large type of crop pests at the cost of slightly more parameters than the baseline EfficientNet. As a result, the proposed scheme can be well facilitated to fully meet the needs of crop pest recognition in smart agriculture.
作者
甘雨
郭庆文
王春桃
梁炜健
肖德琴
吴惠粦
Gan Yu;Guo Qingwen;Wang Chuntao;Liang Weijian;Xiao Deqin;Wu Huilin(College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642,China;Guangdong Provincial Key Laboratory of Agricultural Artificial Intelligence(GDKL-AAI),Guangzhou 510642,China;Guangzhou Key Laboratory of Intelligent Agriculture,Guangzhou 510642,China;National S&T Information Center for Modern Agricultural Industry,Guangzhou 510520,China)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2022年第1期203-211,共9页
Transactions of the Chinese Society of Agricultural Engineering
基金
广东省重点领域研发计划(2019B020214002)
广东省科技计划项目“广东省农业人工智能重点实验室(2021年度)”(2021B1212040009)
广州市科技计划项目(201902010081)。
