A Novel Method for Classification of Butterfly Species Using CNN

Abstract

Researchers studying in the field of lepidopterology want to learn about the family life of butterflies and to examine detailed information such as the shape and species of butterflies. It is of great importance to classify butterflies by invasive methods without harming them. In engineering approaches, the development of reliable, fast and cost-effective systems is suitable to offer solutions to vital problems. In this study, automatic species classification of butterflies is provided by data augmentation with a new method based on Convolutional Neural Network (CNN) for automatic examination and classification of butterflies. In the article, a CNN model that provides automatic classification of 832 butterfly images belonging to 10 butterfly species is proposed. Data augmentation of the proposed model was performed between classes with unbalanced data distribution. To evaluate the effect of the data augmentation process on the performance, the classification process was performed without any data augmentation process. As a result of the data augmentation process, the proposed CNN model reached 93.41% validation accuracy. The proposed CNN model, which does not apply any data augmentation process, has reached 91.72% validation accuracy. The proposed CNN model, which is flexible and highly capable, caused 1.69% performance difference. It is seen that the approach that performs close to or superior to similar studies in the literature is successful. The proposed CNN model is important in that it is both a lightweight and faster system than any pre-training transfer method.