Classifying Overlapping American Sign Language Letters

Juhyoung Lee

jl840 @ duke. edu

Paper PDF

I explored the capacity of a high performing convolutional neural network (CNN) to correctly classify overlapped images of American Sign Language letters. I also modeled an aperture to further investigate how training and performance are affected. I looked at no aperture, a 50 pixel radius aperture, and a 75 pixel radius aperture. Comparing results, I foundnd that adding an aperture lowers accuracy by over 5% and increases odds of overfitting the data.

Paper: Paper PDF

Code and Data: My Code: colab file Link to Kaggle data set: data set