The identification of butterfly families using content-based image retrieval

There is increasing interest in the automatic identification of insect species from images. Here content-based image retrieval (CBIR) is applied because of its capacity for mass processing and operability. A series of shape, colour and texture features was developed that draw on CBIR and allow the identification of butterfly images to the taxonomic scale of family. In our test the accuracy of Papilionidae reached 84% indicating that CBIR is suitable for the identification of butterflies at the family level. Furthermore, experiments with different features, feature weights and similarity matching algorithms were compared. Testing revealed that data attributes such as species diversity, image quality and resolution affected system success the most, followed by features and match algorithms; shape features are more important than colour or texture features in the identification of butterfly families. These findings are important to future improvements in this technology and its applicability. ► An on-line system based on CBIR for butterfly identification is realized. ► Quality of template data is the primary factor affecting identification result. ► Features of shape are more important than texture or colour for classifying butterfly families.