ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (8): 1686-1704.doi: 10.7544/issn1000-1239.2021.20210283

所属专题: 2021人工智能前沿进展专题

• 人工智能 • 上一篇    下一篇



  1. 1(陕西师范大学计算机科学学院 西安 710119);2(陕西师范大学生命科学学院 西安 710119) (
  • 出版日期: 2021-08-01
  • 基金资助: 

Butterfly Species Identification from Natural Environment Based on Improved RetinaNet

Xie Juanying1, Lu Yinyuan1, Kong Weixuan1, Xu Shengquan2   

  1. 1(School of Computer Science, Shaanxi Normal University, Xi’an 710119);2(College of Life Sciences, Shaanxi Normal University, Xi’an 710119)
  • Online: 2021-08-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (62076159, 61673251, 12031010), the Fundamental Research Funds for the Central Universities (GK202105003), and the Innovation Funds of Graduate Programs at Shaanxi Normal University(2016CSY009, 2018TS078).

摘要: 蝴蝶是一种对栖息地敏感的昆虫,自然环境中的蝴蝶种类分布反映了区域生态系统平衡和生物多样性.专家鉴别蝴蝶种类耗时耗力,计算机视觉技术为野外环境中蝴蝶种类自动识别提供了可能.针对野外环境下的蝴蝶图像特征,提出2种新的硬注意力机制,DSEA(direct squeeze-and-excitation with global average pooling)和DSEM(direct squeeze-and-excitation with global max pooling),改进经典目标检测算法RetinaNet,并引入可变形卷积增强RetinaNet对蝴蝶形变的建模能力,实现野外环境下蝴蝶种类自动识别.以mAP(mean average precision)目标检测指标评价模型性能,通过实验结果可视化,分析影响模型性能的关键因素.实验结果显示,提出的改进RetinaNet对自然环境下的蝴蝶识别任务具有很不错的效果, 特别是基于DSEM的RetinaNet;分布平衡的训练集可以提升提出模型的泛化性能;样本的结构相异性是影响模型性能的关键因素.

关键词: 蝴蝶检测, 蝴蝶识别, 注意力机制, 可变形卷积, RetinaNet

Abstract: Butterfly is a kind of insects that are sensitive to the habitat. The distribution of butterfly species in natural environment reflects the balance of regional ecosystem and the biodiversity of the region. To identify the species of butterflies manually is a heavy time consuming work for experts. Computer vision technology makes it possible to automatically identify butterfly species. This paper focuses on identifying the butterfly species via images taken in natural environment. This is a very challenging task because the butterfly wings in the images are always folded and the features identifying the butterfly species cannot be seen. Therefore two new attention mechanisms, referred to as DSEA (direct squeeze-and-excitation with global average pooling) and DSEM (direct squeeze-and-excitation with global max pooling), are proposed in this paper to advance the classical object detection algorithm RetinaNet. And the deformable convolution is simultaneously introduced to enhance the power of RetinaNet to simulate the butterfly deformation in images from natural environment, so as to realize the automatic butterfly species identification task according to the features of butterfly images from natural environment. The very famous criterion mAP (mean average precision) for target detection is taken to value the proposed model, and the visualization is adopted to investigate the primary factors influencing the performance of the predictive model. Extensive experiments demonstrate that the improved RetinaNet is valid in identifying the butterfly species from images taken in the natural environment, especially the RetinaNet embedded with DSEM module. The balanced data can improve the generalization of the predictive model, and the structural dissimilarity of samples is a key factor affecting the performance of the predictive model.

Key words: butterfly detection, butterfly identification, attention mechanism, deformable convolution, RetinaNet