Butterfly family detection and identification using convolutional neural network for lepidopterology

• Published in: International Journal of Recent Technology and Engineering
• Main Author: Bakri B.A.; Ahmad Z.; Hatim S.M.
• Format: Article
• Language: English
• Published: Blue Eyes Intelligence Engineering and Sciences Publication 2019
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85074425452&doi=10.35940%2fijrte.B1099.0982S1119&partnerID=40&md5=63cb688b7111ac525212e32cffc27b4b

Lepidopterology is a branch of entomology concerning the scientific study of moths and the three superfamilies of butterflies. The project aims to help biology students in identifying butterfly without harming the insect. In the studies of lepidopterology, the students normally need to capture the butterflies with nets and dissect the insect to identify its family types. Computer vision is a study on how computers can be used to make high-level comprehension from the input of digital image and videos. By utilizing the latest Image Processing technique, it can identify the correct species of butterfly with high accuracy by using layers of node in a Convolutional Neural Network (CNN). The work process starts with data acquisition (mining the butterfly image automatically from google image search), pre-processing (converting image format and rotation), analyzing and understanding digital images (group images into folders), and to make assumptions of the high complication data from the real world in the process of producing numerical information that can be comprehend by machines in order to form conclusions. Benefits of using CNN is to reduce the need for human and physical intervention in identifying each of the butterfly characters. This makes it easier to expand the database in the future. The image is acquired using Fatkun Batch Downloader to download large number of images. The project is develop using Tensorflow in Ubuntu operating system and interface is in HTML connected to the Python script via Flask. The results of the experiment show that CNN can identify with 92.7 percent of final accuracy with learning saturation (overfitting) of 500 cycle. While testing results shows 62.5 percent of accuracy in predicting new datasets. © BEIESP.