1

I used Azure Cognitive Vision API to extract the text from a cheque image. But, it is not correctly extracting the text from cheque. Example MICR code having characters like " || are incorrectly read into some other digits. Is there any way we can work on to improve the accuracy or set some context to specifically extract text from cheque images?

Sample Image used for Vision API: enter image description here

Output of Vision API:

MDBS
Date
2 16
NOT NEG
AS PAYES
D
D
M
M
Y
Y
Pay
Hexagon metrology Asia Pacific
or Bearer
SECURA 3011678/2 06/19
Singapore Siahd hundred and niudy- S$ 896<
Dollars
Sit only
2M PRECISION ENGRG
DBS Bank Ltd
PTE LTD
Cheque No.
Bank/Branch Code
Account No.
Please sign above this line
1;91 308 2081717 00540 0549001996
atulsri
  • 33
  • 1
  • 7
  • please add more details about the sample image and output that you are getting. – Ram Feb 16 '21 at 05:36
  • Ram, I added the sample image I used and its output produced by Vision API. I blackened some text due to security concerns. – atulsri Feb 16 '21 at 07:10

2 Answers2

1

This custom tags is currently not available. However, our engineers are working to bring this functionality to Computer Vision. You can use the custom vision to detect.

Azure Form Recognizer does a fantastic job in creating a viable solution with just five sample documents. It performs end-to-end Optical Character Recognition (OCR) on handwritten as well as digital documents with an amazing accuracy score and in just three seconds.

If you are extracting only text, tables and selection marks from documents you should use layout, if you also need to extract key value pairs you can train a custom model or use a pre-built (Invoice, Receipts, Business Cards). Layout results (text, tables and selection marks) are included in all the Analyze outputs (custom and pre-built) in the readResults (text) and pageResults (tables) of the JSON output.

• Layout – extract text, tables selection marks no training required

• Pre-built – Invoice, Receipts, Business Cards – extract values of interest from these type of documents

• Custom – Extract key value pairs trained on your own documents

Here is the snapshot of output text computer vision API. enter image description here.

Please follow the Computer Vision API Frequently Asked Questions that could help.

Ram
  • 2,459
  • 1
  • 7
  • 14
  • 1
    Thanks Ram for answering. I will explore this custom service. – atulsri Feb 16 '21 at 07:00
  • Ram, custom vision will not fulfill my purpose.I need to extract data from cheque image. I used both bitonal and gray images, but no API call returned successful result accurately. – atulsri Feb 16 '21 at 11:40
  • Thanks for the details. Azure Form Recognizer does a fantastic job by using the custom model, you can see this study for performance. https://cazton.com/blogs/executive/form-recognition-azure-aws-gcp – Ram Feb 16 '21 at 12:41
  • Hi Ram, Thanks for giving one more option. However, I already looked at Form recognizer but not in budget when you have thousands of documents. It charges are around 50$ per 1000 documents for custom layout. – atulsri Feb 17 '21 at 06:13
0

Not just Azure, even Google cloud Vision API or AWS Rekognition Text Detection has the same issues while reading the MICR code. The separators are predicted as either 1 or : (colon). We've built a custom model to overcome such issues. https://api.arya.ai/cheque-extraction

Deeks
  • 1
  • 1