基于细粒度特征增强交互网络的植物病虫害识别  

Plant Pest and Disease Identification Based on Fine-Grained Feature Enhancement Interaction Network

作　　者：杨筝[1] 程云志[2] 常开心 侯彦东[3] 陈政权 YANG Zheng;CHENG Yunzhi;CHANG Kaixin;HOU Yandong;CHEN Zhengquan(School of Electrical Engineering,Yellow River Conservancy Technical Institute,Henan Kaifeng 475000,China;School of Computer and Information Engineering,Henan University,Henan Kaifeng 475004,China;School of Artificial Intelligence,Henan University,Zhengzhou 450000,China)

机构地区：[1]黄河水利职业技术学院电气工程学院,河南开封475000 [2]河南大学计算机与信息工程学院,河南开封475004 [3]河南大学人工智能学院,郑州450000

出　　处：《河南大学学报(自然科学版)》2024年第6期722-729,共8页Journal of Henan University:Natural Science

基　　金：国家自然科学基金资助项目(61374134);河南省自然科学基金资助项目(232300421149)

摘　　要：针对同种植物不同病虫害特征差异小而不易识别的问题,本文提出了一种细粒度特征增强交互网络(FFE-HBP).首先,利用不同卷积层的互补优势来减少特征信息在提取过程中的损失.然后,通过细粒度特征增强模块降低特征粒度并从细粒度层面上增强特征的显著性.最后,采用分层双线性池化模块将增强后的多层特征进行交互,以捕捉各层之间的特征关系,同时集成多个跨层交互信息提高模型对判别性特征的提取能力.实验结果表明,在番茄数据集上具有明显的优势,其识别率达到93.95%.此外,在苹果、玉米和葡萄的数据集上验证了FFE-HBP的有效性和泛化性.该模型能够有效提取判别性的细粒度特征,提高对相似病虫害的识别能力.For the problem that it is not easy to identify different pest and disease characteristics of the same plant with small differences, this paper proposes a fine-grained feature enhancement interaction network(FFE-HBP). First, the complementary advantages of different convolutional layers are exploited to reduce the loss of feature information in the extraction process. Then, the feature granularity is reduced and feature saliency is enhanced from the fine-grained level by a fine-grained feature enhancement module. Finally, the enhanced multi-layer features are interacted using a hierarchical bilinear pooling module to capture the feature relationships between the layers, while integrating multiple cross-layer interaction information to enhance the model's ability to extract discriminative features. The experimental results show a significant advantage on the tomato dataset, with a recognition rate of 93.95%. In addition, the effectiveness and generalization of FFE-HBP are verified on the datasets of apple, corn and grape. The model can effectively extract discriminative fine-grained features to improve the identification of similar pests and diseases.

关 键 词：病虫害识别 深度学习 细粒度识别 神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]