Deep learning based butterfly species identification system through wings pattern
Loading...
Date
2017-06
Authors
Tan, Yan Jia
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
For entomologists, butterflies act as a bio-indicator for the health of a forest or an ecosystem. From the observation of butterflies, the condition of the forest can be determined and counter-measure can be taken earlier to neutralize or minimize the threat. The automated identification application, is helpful in assisting entomologist to survey an area with the implementation of Machine Learning (ML). It was once a difficult challenge due to uncertainties in the captured images such as obstructions and high variations of pose. With the hardware getting better and powerful, the concept of Convolutional Neural Network (CNN) that was once an idea can be realised. It is evident in the object recognition challenge organised by ImageNet, ImageNet Large Scale Visual Recognition Challenge (ILSVRC), where CNN produced state-of-the-art result. In this project, Inception V3 is used as the core engine to power the species recognition system that aims to help the entomologists to identify the butterfly species. Optimisation is performed by experimenting in stages with several training parameters to obtain the best value for this unique purpose. Using transfer learning, the best result recorded is 14.3% error with capability of identifying ten species, which includes two pairs of butterfly species with similar features. With Graphics Processing Unit (GPU), the processing time is less than 0.15 seconds per image which is three times faster than using a Central Processing Unit (CPU).