3
$\begingroup$

I am trying to implement an text detector using MaskRCNN such that the model detects the label and value as shown in the image below.

Detecting the same is easier for fields like page date and order no. but fails for invoice number and salesman shown in image below. Is there any pretrained models for the same? or some other technique that might help.

$\endgroup$
0
$\begingroup$

Here is a paper doing exactly what you want with detailed explanation of architecture.

There is not much sense of me doing a TL;DR here because they have some special labeling of data etc... but the gist of it is LSTM if you do some pre-processing steps and get your data in right format.

$\endgroup$
3
  • $\begingroup$ Solution I am designing is for scanned invoices in image format. $\endgroup$ – hR 312 Dec 26 '19 at 9:26
  • $\begingroup$ and the difference is? $\endgroup$ – vienna_kaggling Dec 26 '19 at 9:32
  • 1
    $\begingroup$ I'll have to use a pretrained OCR to read the text and what I want is my model to readimage given above as Invoice No. 6 but pretrained OCR might fail in doing so. $\endgroup$ – hR 312 Dec 26 '19 at 9:37
0
$\begingroup$

You can use CTPN to localise text and later use OCR for capturing the text , refer to the git page it will serve your purpose .you can pretrain on your custom data set as well.

$\endgroup$
1
  • $\begingroup$ Thanks for your input. May I know what's wrong in using MaskRCNN as CTPN is an implementation of Faster RCNN itself? $\endgroup$ – hR 312 Jan 2 '20 at 9:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.