Case Study

How we identified sea life
using deep learning

Intro

During my last holiday I was Scooba Diving with one of my best friends Sven. He is an experienced Scooba Diver unlike myself. During our underwater exploration we were able to witness amazing wilflife. We even saw a seaturtle. But Sven also saved me from a lot of potentially dangerous fish. As he later explained I was getting too close to a puffer fish that has neurotoxins and could be deadly.

While looking at the picture of our trip he pointed out the different species of fish and their characteristics. Then I thought, how cool would it be to create a deep neural network to classify these fish?

INDUSTRY
PythonDeep LearningArtificial Intelligence
Tech used

Tensorflow, Python, Tensorflow Hub

Result

The fish detection algorithm could classify twelve sea life categories with 90% precision.
The accuracy could have been better but because of lacking data it was difficult to get a better result.

Details

It turned out the fish detection algorithim was much more complicated than expected. I had to deal with low quality photos, poor lightning conditions, objects blocking the sight, divers in the photo... But the biggest hurdle was getting the data. Unfortunately, I did not find an online fish database I could have used. So preparing the training data was done manually.

I retrained one of the tensorhub modules to classify the fish to make up for the missing data. That turned out quite well.

The images were resized into one standard format during pre-processing.

Let’s build the future together

Test a business idea with your target audience or show its potential to your investors. Proprietary tools and artificial intelligence – making software development easy and transparent.