Thursday January 25th, 4:30-5:30pm in POST 126
What’s That Plant?
WTPlant is a Deep Learning Framework to Identify Plants in Natural Images
An interdisciplinary area encompassing Botany and Computer Science presents the challenging task of automating the identification of plant species. A highly accurate automated system may solve the botanical taxonomy gap and be used to identify new species, control the balance of ecosystems, and several agricultural activities. This is a very difficult problem due to the complexity of natural image processing by automated computer vision techniques. Images with different objects and/or plants on the background impose heavy complications to the problem. Multi-scale images and other variants such as position and inclination also increase the complexity and have to be addressed carefully. The plant itself presents multiple organs (leaf, flower, fruit, bark, etc.) with different characteristics throughout the seasons of the year. A large number of species and the similarity between them add further obstacles to the proposed framework. By creating customized Convolutional Neural Networks with Residual Blocks to extract deep features and Inception Modules to handle the multi-scale issue, the proposed framework seeks to develop a robust and highly discriminative deep learning method for this image classification problem. Initially focusing on Hawaiian plants, this proposal details how to extract representative training samples from annotated natural images and presents the creation of a benchmark. The objective is to produce the most accurate system to classify natural images of Hawaiian plants and make it available online for botanists, gardeners, tourists, and the entire community to use it across the Hawaiian Islands.