부산대학교 도서관

In the study, we describe a complete system for butterfly image classifier with the ResNet 18 architecture as the backbone structure. The very colourful wings of the butterflies, as well as their patterned structures, are faculties that make the image classification of butterflies not an easy task for image classification algorithms. Our approach, entitled “two-step methodology to deal with the challenge” is as follows: We first make the ResNet18 architecture faster to suit more contributing the specific characteristics of butterflies by the addition of techniques like attention and spatial transformer networks. The goal of this change is for the network to get more sensitive to the features at varying levels which is key for the correct classification of the image. Firstly, our parrot model undergoes fine-tuning, which works to adjust the already trained ResNet 18 model for the butterfly classification task. By taking advantage of pre-trained networks on the big dataset of butterfly pictures, we plan to make use of the extracted features to reach an optimal performance level for the specific task. Our findings prove that our model is more accurate than any currently existing approach, achieving state-of-the-art results on the butterfly image databases for global benchmarking. Besides that, we have the network of such representations and feature activations reconstructed which gives us a glimpse into the important butterfly classification features. In conclusion, our study is a starting point for a deep learning approach to butterfly image classification; it will enable keeping tabs on the biodiversity of a particular location, contribute to the conservation efforts of endangered species, and promote ecological research.