Money Recognizer

Inspiration

One day, we were brain-storming for a side project with my friend. She told me that she has a visually impaired relative and he has some problems to recognize banknote bills. So as I am an iOS/ML developer, I decided to create an app that will solve this problem with computer vision and deep learning.

What it does

You show the money to iPhone camera and it tells you the denomination.

How I built it

I created a dataset and took 120 pics of each denomination. I upload this dataset to CustomVision. On CustomVision I trained a fast model and then advanced model in under an hour. After tranining finished, I tested the model with the images it hasn't seen before. And then when I am satisfied with the accuracy of the model, I exported it as a CoreML file. I used a sample application that CustomVision shared on Github. Changed the code to use my model and add it a text to speech.

Challenges I ran into

Normally I train my models using Keras. When I try to train a model with my GPU, it takes many hours to reach this level of accuracy. And I need to write some code to convert the Keras model to CoreML model. With CustomVision, It is very easy and it takes much less time.

Accomplishments that I'm proud of

I built an app which can recognize money from images under an hour.

What I learned

I learned we can export many models (Tensorflow, CoreML, Onnx etc.) from CustomVision an even publish an api for our model.

What's next for Money Recognizer

As I see CustomVision offers export option as Tensorflow for Android apps. I guess I will build an Android app using this model.

Built With

• azurecognitiveservices
• coreml
• customvision
• swift