I want to build an open Source OCR just like MathPix. There is already a model to extract LaTex from the image by Harverd NLP's im2markup but the problem is that their data has been trained and tested only for the Cropped Portion of Latex Only. Somehow, I might be able to implement the code using their idea but the problem is that I have hybrid images of text as well as mathematical equation.
My idea is that if I am able to find the portion/bounding box/ region where the equation is present, I can crop and then apply Harvard's model to get the latex. But the main problem is that How am I supposed to find the area of interest where equations are present??
All the object recognition algorithms like
YOLO, Fast,Faster RCNN, MTCNN etc won't work here in this case as I am dealing with the text within text than any object in text and vice versa.
Could someone please suggest a method hoe could I get the area where I can get te bounding box of equation and crop the image to get the LaTex from that image using