Multiple butterfly recognition based on deep residual learning and image analysis

• Published in: Entomological research 2022-01, Vol.52 (1), p.44-53
• Main Authors: Xi, Tianyu, Wang, Jiangning, Han, Yan, Lin, Congtian, Ji, Liqiang
• Format: Article
• Language: English
• Subjects: Accuracy; butterfly identification; Datasets; Deep learning; deep residual learning; Image processing; image recognition; multiple species recognition
• ISSN: 1738-2297; 1748-5967
• DOI: 10.1111/1748-5967.12564

Insect recognition is crucial for taxonomy. It helps researchers to process tremendous and various ecology data. Most studies focus on fine‐tuning the deep learning network or altering the algorithm to enhance the identification accuracy, and some useful tools have been generated with these methods. This study focuses on the influence of image data on the recognition model. The single data set source of the existing automated identification tools is relatively simple, and the competition‐based data set released only focuses on evaluating the model at present. For the first time, this article integrates butterfly image data sets from multiple sources, covered illustrated books, and popular butterfly science websites. The image types include standard specimen images, illustrated book scan images and camera shots. In addition, these images included not only fixed poses, but also various other images of butterflies in natural poses. The size of these images is also various. The testing data set is new data that does not belong to the training set, which also verifies the generalizability of the model, indicating that in practical applications this model can identify new images. This testing method is a breakthrough compared to the previous work. We designed different data sets using the ResNet18 network to train a classifier, which achieves a validation accuracy of 86% in the end of the analysis. By adjusting the data sets, the accuracy changes as well. This study provides a method to recognize hundreds of butterfly species and analyzes the testing progress from the point of view of data. It is the first to combine butterflies from multiple countries in a single data set, with a recognition accuracy that outperforms previous experiments, to the best of our knowledge. We further analyze the testing results of butterfly recognition at the family and genus level. We perform two more experiments to demonstrate the model in the case of similar species or genus.