PixelLink

What it does

Given an Image identify all instances of text present in it. This is useful for autonoumous cars that can accurately detect road signs. This is also useful for people with visual impairment that need help in reading signs and signals in and around their environment. The algorithm does this fast enough that a 60fps video can be recognized in real time. >90% accurracy.

How I built it

Rather than using the general semantic segmentation which introduces several challenges due to different charachters eing close to each other, I tried to implement a concept called Instance Segmentation. Tools used are python, tensorflow and opencv.

Challenges I ran into

It was difficult to understand certain parts of implementation of the paper which needed to be tweaked.

What I learned

What Instance segmentation is and how to implement it in code.

What's next for PixelLink

Test it with a robust dataset and maintain its accuracy.

Built With

• python
• tensforlow