Fine-Grained Visual Classification for Plant Analysis

Ram J. Zaveri*, Matthew Keaton, Meghana Kovur, Cole Henderson, Gianfranco Doretto, Ph.D.
Lane Department of Computer Sciences and Eletrical Engineering, West Virginia University, Morgantown, WV 26506

Presentation Category: Science & Techonology (Oral Presentation)

Student’s Major: Computer Science and Neuroscience

Object classification from images is a standard problem in computer vision. State of the art techniques are based on deep learning, a subfield of machine learning. While they perform well for generic object classification such as determining whether there is a cat or a dog in an image, there are application domains where they still fall short. One of them is the automated classification of plant species. This is a case where different plants might sometimes look very similar in pictures, while the same plant can appear very differently, because of the concurrent effects involved in the image formation process, which involve the shape of the scene, its material properties, the illumination conditions, and the viewpoint. We regard this scenario as a fine-grained visual classification problem. To address it, we are curating a new large-scale dataset. We developed a tool for annotating plant images, which we use for labeling the visible plant organs. The dataset will serve as a benchmark for training different fine-grained visual classification techniques based on deep learning and establish their potential for image species classification.

Funding: National Science Foundation (NSF)

Program/mechanism supporting research/creative efforts: WVU's Research Apprenticeship Program (RAP) & accompanying HONR 297-level course