Leaves

Project Overview

Humans have the ability to easily tell apart different animal and plant species. On the other hand, it is more difficult for computers to differentiate between different species. The goal of our leaves or “Unbe-leaf-able AI” project was to see if a computer can use neural networks and image recognition to correctly identify 185 different species of leaves via feature extraction. We tested the Resnet18 and VGG11 models to find which model has the greatest accuracy in predicting the correct leaf species.

Project Question

Are we able to classify different species of leaves using Machine Learning?

Which network architecture would work the best to differentiate between leaf species and to what accuracy?

Action Steps

We first defined and used a Dataloader class to load data from the Leafsnap dataset. After we gathered our leaves data, we trained the Renset18 and VGG11 networks to determine which model was more accuracy in differentiating between different leaf species. After determining the initial accuracy of both networks, we re-ran them over 5 Epochs to further increase the accuracy of each neural network. We also compared the pretrained version of each network to the non-pretrained version and found that the pretrained version with transfer learning produced more accurate results.

Results