PirateNet

Ship detected
No ship detected
Sonar vs Camera

Inspiration

We’ve always had a passion for machine learning, with PirateNet being our first full application of our knowledge and skills. The pirate theme was very fitting as it forced us to think more creatively and how we would come up with projects related to it. We’ve seen other students be able to create amazing things with machine learning and we wanted to participate in this hackathon and go for this specific project.

What it does

Our machine learning model detects objects in the ocean via sonar images. Using TensorFlow, Pickle, and numpy, we turned our dataset into grayscale, 1D vectors. We had three hidden layers, each using the ReLu activation function. We also used a dropout of 0.2 to make sure we did not overfit the data. It then runs through 25 epochs and displays the final accuracy of our model.

How we built it

The first thing that we did was get a suitable dataset. We made sure that there was relevant training data that we used. We decided to use a dataset of sonar images. After that we created a basic deep neural network through TensorFlow. We wanted to expand on this network, but it became too computationally expensive with the images we were using.

Challenges we ran into

Finding a dataset was difficult. We originally looked on Kaggle for our data but there was little we could find. It was made especially difficult because during the Hackathon we weren’t able to download a lot of datasets because of the slow download speeds. But, we were able to find a good dataset that features shipwreck images and sonar scans. Additionally, we could not use some python packages due to errors with installing them.

Accomplishments that we're proud of

We achieved a high accuracy of over 95% and we successfully built our own model.

What we learned

It takes time for a project to fully get started. We bounced through many different ideas and iterations before arriving at what we have. There were also many obstacles we faced such as identifying a proper data set or using specific imports. We had to research and look into how to use specific things and that allowed us to steadily make progress.

What's next for PirateNet

We have plans to expand PirateNet to identify specific objects. Currently, it can identify underwater wreckage from sonar scans. However, further research will allow us to train our model to identify things like airplanes or other objects lost in the ocean.