OCR detection gets worse with cropped images
#49
by
Asdmkasomo
- opened
Hi All, asking for some help here -- we are trying to recognise text from a document but cropping of the document would result in GOT OCR being unable to detect the words.
However, it is able to detect if it was with the original image.
Able to know what should be the re-sizing be to match with the model.?
Think about it this way: when you crop the image, the number of pixels is reduced. By the time it goes through the model’s feature extraction, some features are lost, which makes detection fail. You should try resizing the cropped image first (even using nearest-neighbor resizing is fine). Then the model will be able to detect it, since it has already been trained on characters of various sizes.