Organ-based Visual Classification for Plant Analysis

Cole Henderson*, Meghana Kovur, Ram Jitendrabhai Zaveri*, Matthew Keaton, and Gianfranco Doretto

Lane Department of Computer Science and Electrical Engineering, West Virginia University, Morgantown, WV 26506

Presentation Category: Physical Sciences & Engineering (Poster Presentation #149)

Student’s Major: Computer Science and Computer Engineering

The great amount of biodiversity on our planet makes it difficult to have complete, accurate datasets that are readily available to experts in the field. Often, even if an expert is knowledgeable about a certain species, it can be difficult to identify samples with a high degree of accuracy. Current computer vision approaches allow for repeated, accurate identification of species, but only from specially prepared samples, like herbarium sheets. They often use attention-based feature learning, which works well on most datasets, but loses effectiveness when applied to plants. To be truly useful, an approach needs to be able to handle data collected from the wild in varying levels of quality and distribution of features. Our approach attempts to solve this issue by using an object detector to direct attention-based learning. We take an image of a plant, then break it into its subsequent organs using a detector. The individual organs are then sent to one of a host of species classifiers, each of which is specialized to classify the plant species given a particular organ. These predictions are then fused together to generate the final output prediction. We have developed a neural network species classifier based on a dataset of images that we have curated and are currently evaluating different techniques for optimizing its prediction accuracy.

Funding: National Science Foundation

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