This project was the last one before I entered university. I like it because it was the first time that I solved a problem using deep learning. The idea came to me when I first saw the Seeing AI app from Microsoft.
It had many features that could help a person with visual disabilities and ease their life. One interesting feature was to identify currency bills when paying with cash. I tried to implement this for the people of my country because the original app only supported hard foreign cash.
As there was no dataset to train with, I spent one month collecting a good one, and meanwhile, I was studying possible approaches to this kind of problem. I decided to use transfer learning on MobileNet because I wanted to run the model on an android device. I realized TensorFlow has an excellent framework for on-device inference. So, I started to explore the documentation, and after a couple of weeks, I came up with this result:
Even though I did not fully understand each tiny detail in the model that I trained, I still learned a lot of things that I did not know before starting the project. For me, the most impressive part of this project was the fact I did not use any hand-crafted feature similar to what I used in my advanced line follower robot and shooting simulator game.
I knew there was still a lot more to explore in this field, and I had barely scratched the surface of it!
This project later motivated me to learn more about machine learning and deep learning.