Personal Experience
#13
by fluxnad - opened
Personal thought. PaddleOCR-VL does a great job on text recognition. What I noticed in complex tables is a cell detection issue. When the table relies on alignment and spacing instead of clear cell borders, the model sometimes merges cells or assigns values to the wrong column. In my example, subtotal rows like “S/Total” lose the correct column alignment, and at times a whole column region gets treated as one cell when the structure is not clearly labeled and it droped the values of 2020
this a part of a table
and this is the output of the ocr
heres another exemple
and heres the output
and another one:
output:
by the way i set it the use_doc_unwarping=False to refer to the original image
