This started as a project last summar. Persian language is really complicated and open-source OCR applications for our language are limited. I thought maybe using graphs might be the solution. I started by watching this great Stanford University course about graphs[https://www.youtube.com/watch?v=JAB_plj2rbA&list=PLoROMvodv4rPLKxIpqhjhPgdQy7imNkDn]. I first ran a classification test on single Letters with a really small dataset. To improve the model I found the paper[https://www.sciencedirect.com/science/article/abs/pii/S003132031400274X] that used similarity learning with Graph Edit Distances and so I implemented it with the help of this git repository[https://github.com/priba]. This project is still underwork and not complete !
First the image is skeletonized and and then a graph is formed from the character each node being a keypoint of the letter and node values being normalized x and y of the points in the image.
|
Data Sample |
|
Skeletonized |
|
Graph |