HealthRice

A web app that uses machine learning to diagnose Hispa and BrownSpot on rice leaves.

Comment

Inspiration

Do you know what the most popular and crucial food staple in the world? If you guessed rice from our name, you would be correct. More than 3.5 billion people depend on rice as their primary food source. However, rice can get diseases that are extremely hard to diagnose just by looking. Popular diseases among rice include Brownspot and Hispa. When people eat diseased rice, they can often times get extremely sick. This is a huge problem in underdeveloped countries where rice is the main food source and there is not much access to hospitals. Luckily, HealthRice is here to help.

What it does

Healthrice is a web app that uses the power of machine learning in order to diagnose images of rice leaves that may or may not be contaminated with either Hispa or BrownSpot.The UI is extremely intuitive for people to be able to easily get their results. Afterwards, if desired, the user can input their phone number to be reminded of the rice’s condition via an SMS. The machine learning model is able to have a training accuracy of 90% while it is still able to diagnose the image quickly.

How we built it

Healthrice was made using Tensorflow, Twilio, and Streamlit. Tensorflow was used the create the powerful Convolutional Neural Network that is behind HealthRice. Next, the Twilio API was used to send a SMS to the user’s phone. Finally, Streamlit was used to create the frontend of HealthRice. Streamlit allowed are UI to be simple but beautiful.

Challenges we ran into

Some mentors during this Hackathon can all agree with me that I was having trouble with my machine learning model file size and Github. Github has a file size limit that my saved machine learning model exceeded. So I tried using Tensorflow Lite only to find out that it was still too large. I soon solved this problem by removing a couple of training images. This somehow boosted the training accuracy up to 90 percent and kept the file size low even without Tensorflow Lite. Because I had two different areas of this project, the machine learning model and the frontend, I tried to keep things in two different directories. This led to untracked errors in Git and Github which forced me to create a new repo. (This is why commit history looks weird). However, with the help of a few mentors, the problem was solved!

Accomplishments that we're proud of

I'm super proud of being able to make a CNN for the first time and actually deploy it to a web application. Our accuracy was quite high during training and I am super proud of that.

What we learned

We learned that it might sometimes be a bad idea to try and stuff everything into a Github repo. For my training data, we had a lot of images so I couldn't upload them to Github due to file count limits. However, with the help of a mentor, I was able to put them on a gitignore file and things went smoothly again.

What's next for HealthRice

In the future, we plan on adding more classes to our machine learning model. For now, our model can diagnose 3 classes: Healthy, Brownspot, and Hispa. We plan on adding more diseases that our model can diagnose.

Built With

Share this project:

Updates