基于注意力卷积增强特征网络的昆虫图像识别  

Image recognition of insects based on attentional convolution enhancement features networks

作　　者：韩巧玲[1,2,3] 周晗 赵玥[1,2,3] 王禹沣[1,2,3] 赵燕东 王海兰[1] HAN Qiao-ling;ZHOU Han;ZHAO Yue;WANG Yu-feng;ZHAO Yan-dong;WANG Hai-lan(School of Technology,Beijing Forestry University,Beijing 100083,China;Key Lab of State Forestry Administration for Forestry Equipment and Automation,Beijing Forestry University,Beijing 100083,China;Beijing Laboratory of Urban and Rural Ecological Environment,Beijing Forestry University,Beijing 100083,China)

机构地区：[1]北京林业大学工学院,北京100083 [2]北京林业大学林业装备与自动化国家林业局重点实验室,北京100083 [3]北京林业大学城乡生态环境北京实验室,北京100083

出　　处：《计算机工程与设计》2025年第4期1072-1078,共7页Computer Engineering and Design

基　　金：国家自然科学基金青年科学基金项目(32101590);国家自然科学基金面上基金项目(32071838)。

摘　　要：为解决由于五大连池地区昆虫样本量少、类别分布不均导致昆虫识别准确性低的问题,提出一种基于注意力卷积增强特征的匹配网络(feature-enhanced matching network,FEMNet)。采用随机欠采样对数据集进行平衡处理;通过提出特征上下文嵌入模块,增强昆虫全局和局部特征的提取能力;基于匹配网络实现样本间特征的灵活匹配,提高小样本下昆虫图像识别精度。实验结果表明,对于小样本昆虫数据集,FEMNet方法比次优方法MatchingNet准确率提升4.5%、召回率提升4.8%、精确率提升6.1%、F1值提升5.3%,说明该方法能够准确自动识别昆虫,可为后续昆虫学研究提供技术支持。To solve the problem of low accuracy of insect recognition in Wudalianchi area due to the small sample size and unba-lanced distribution of insect datasets,a feature-enhanced matching network(FEMNet)was proposed.Random undersampling was used to balance the dataset,a feature context embedding module was proposed to enhance the ability to extract the global and local features of insects.The feature matching between samples was implemented flexibly based on the matching network.Experimental results show that for small sample insect data sets,the accuracy rate of FEMNet method is increased by 4.5%,the recall rate is increased by 4.8%,the precision rate is increased by 6.1%,and the F1 score is increased by 5.3%compared with that of the second-best method MatchingNet.This shows that the method can effectively extract and distinguish the subtle features of insects,and accurately and automatically identify insects,which can provide technical supports for the follow-up entomological research.

关 键 词：昆虫识别 图像处理 五大连池 小样本学习 匹配网络 不平衡学习 随机欠采样 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]