Predictive Image Classification

Inspiration

Our team had a deep interest in what capabilities Computer Vision and similar technologies had, and how we could leverage those to automate processes. In this project, we adapted several convolutional neural network (CNN) models to identify images, and then analyzed those classifications to assess their performance.

What it does

It identifies and classifies objects in an image, with an 85% accuracy.

How we built it

We adapted several models into Python, input provided data sets and output predictive classifications. We found that CNNs with more layers had the potential to perform better across various data sets when given the right parameters.

Challenges we ran into

The main challenges we ran into were our results when attempting to classify humans into their respective genders. The results for human image classification provided extremely low accuracy scores for all models, which we decided to omit from our other model statistics as it would greatly skew the overall accuracy scores. However, the ResNet CNN model showed great potential. ResNet's base model started with 50 layers, which gives it a greater ability to be fine tuned to handle more complex features, however deeper models increases the potential for overfitting of the data, which would have to be handled through extensive feature engineering. The issue we faced here was that it would be difficult to fine tune and further train the model in this short of a time frame, so the fine tuning was not performed.

Accomplishments that we're proud of

Thankfully, our team did gather enough results from the provided data sets to establish and strongly support our conclusion. VGG19 turned out to have the greatest True Positive Rate of 85%, which makes it the best model for the provided data set.

What we learned

Limitations that VGG19 faces are that it has a difficult time segmenting objects if they have a dominant background or are just not generally distinguishable. Yet still, it was able to correctly identify objects 85% of the time for our set of nearly 200 data points.

What's next for Predictive Image Classification

We would like to further train ResNet to recognize and classify images more accurately, more specifically to conduct human/gender identification and facial recognition.