Challenge DATA & AI by Ens'IA and Neovision

The challenge:

This challenge was organized by Ens’IA and Neovision in 2021. It lasted a month and a half and was about the well known subject of Optical Character Recognition (OCR).

➡️ The goal was quite simple but still very challenging:
building the best model to recognize the characters in the dataset given by Neovision. (private dataset).

➡️ There were more than 30 participants divided into 12 teams that competed against each other. The best teams were rewarded with 6 Nvidia Jetson cards!

The results:

With my teammates Clément and Valentin, we managed to build an AI that achieved a macro-averaged accuracy of 0.78 in the provided test dataset. It allowed us to finish in second place.
My team managed to end in second position with a macro-averaged accuracy of 0.78.
Learn more

The model:

My model is available in my github french version.
We tried different methods and worked on transfer learning, on image preprocessing, on segmentation, etc… and tried different techniques developed in recent research papers.
The best result used transfer learning with MobileNet and its pre-trained weight on ImageNet