Date of Graduation


Document Type

Problem/Project Report

Degree Type



Statler College of Engineering and Mineral Resources


Lane Department of Computer Science and Electrical Engineering

Committee Chair

Gianfranco Doretto

Committee Member

Katerina Goseva-Popstojanova

Committee Member

Donald Adjeroh


Image-based plant species identification in the wild is a difficult problem for several reasons. First, the input data is subject to a very high degree of variability because it is captured under fully unconstrained conditions. The same plant species may look very different in different images, while different species can often appear very similar, challenging even the recognition skills of human experts in the field. The large intra-class and small inter-class image variability makes this a fine-grained visual classification problem. One way to cope with this variability and to reduce image background noise is to predict species based on the plant organs that are visible. In this work, we designed a set of classifiers that predict species conditioned on the plant organs. Due to the lack of a dataset for testing this task, we curated one ourselves with 1000 species and 5 organ types per species, which is based on the data used in the PlantCLEF 2015 challenge. We designed the bank of organ-based plant species classifiers with convolutional neural networks, and we performed a comparative study that shows how different techniques for improving the classifier training affect the accuracy of the identification for each of the organ type.