基于残差网络的自然环境下蝴蝶种类识别  

Identification of butterfly species in the natural environment based on residual network

作　　者：李飞 赵凯旋 严春雨 闫建伟[3] 邢济春[4] 谢本亮 LI Fei;ZHAO Kai-Xuan;YAN Chun-Yu;YAN Jian-Wei;XING Ji-Chun;XIE Ben-Liang(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;Semiconductor Power Device Reliability Engineering Research Center,Ministry of Education,Guiyang 550025,China;School of Mechanical Engineering,Guizhou University,Guiyang 550025,China;Institute of Entomology,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵阳550025 [2]半导体功率器件可靠性教育部工程研究中心,贵阳550025 [3]贵州大学机械工程学院,贵阳550025 [4]贵州大学昆虫研究所,贵阳550025

出　　处：《昆虫学报》2023年第3期409-418,共10页Acta Entomologica Sinica

基　　金：国家自然科学基金项目(61562009);国家重点研发计划课题(2021YFD110030);贵州科技计划项目(黔科合成果[2019]4279号,黔科合平台人才[2019]5616号);半导体功率器件教育部工程研究中心开放基金项目(ERCMEKFJJ2019-(06))。

摘　　要：【目的】蝴蝶属鳞翅目(Lepidoptera)昆虫,其对生存环境敏感,能够作为区域生态环境的指示物种,自然环境下蝴蝶种类自动识别对生态系统稳定有重要意义。现有研究中蝴蝶种类和数量较少,且多以标本图像作为识别对象,鉴于此,本研究构建了自然环境下蝴蝶图像数据集,提出一种以残差网络为基础的蝴蝶种类识别模型LDResNet。【方法】首先,引入可变形卷积,增强网络对不同形状蝴蝶图像的特征提取能力,获得更细粒度的特征;其次,在可变形卷积后嵌入注意力机制,增大蝴蝶特征权重,降低冗余信息干扰;最后,利用改进的深度可分离卷积降低模型参数量。【结果】在自建数据集上实验,LDResNet模型取得了87.61%的平均识别准确率,较原始模型提升了3.14%,模型参数量仅为1.04 MB。【结论】LDResNet模型相较其他模型,在平均识别准确率和参数量方面均有明显优势,本研究模型可为自然环境下的蝴蝶种类自动识别提供技术支持。【Aim】Butterflies,as lepidopteran insects,are sensitive to their living environment and can be indicator species of the regional ecological environment.Automatic identification of butterfly species in the natural environment is of great significance to ecosystem stability.In the existing studies,there are few species and numbers of butterflies,and most of them take specimen images as recognition objects.Therefore,the butterfly image data set in the natural environment was constructed,and a butterfly species recognition model LDResNet based on the residual network was proposed in this study.【Methods】Firstly,deformable convolution was introduced to enhance the network feature extraction ability to different shapes of butterfly images,and to obtain more fine-grained features.Secondly,the attention mechanism was embedded after the deformable convolution to increase the weight of butterfly features and reduce the interference of redundant information.Finally,the number of model parameters was reduced using an improved depthwise separable convolution.【Results】Experimenting on a self-built dataset,the LDResNet model achieved the average recognition accuracy of 87.61%,a 3.14%improvement over the original model,with only 1.04 MB of model parameters.【Conclusion】LDResNet has obvious advantages over other models in terms of average recognition accuracy and number of parameters,and this research model can provide technical support for automatic identification of butterfly species in natural environments.

关 键 词：蝴蝶 残差网络 可变形卷积 注意力机制 深度可分离卷积 

分 类 号：Q968[生物学—昆虫学]