problems with data detection with Invoice prebuilt model

[YanysM]

Hello, 

I'm currently  working on automating invoice porcessing using an Ai prebuilt model.

While it genrally works well with most invoices, i've noticed ocasional issues with detecting all items table properly.

I can understand the importance of the invoices quality or certain formats can be a challenge for the AI Models as it can confuse the prebuilt model, but it can correctly  recognize most of the items within the invoices, but occasionally it just skips some of them or the data detection is wrong.

Sometimes, reprocesing the invoice can fix the issue, but thats not the ideal solution.

 I've also tried to train a model, but i have the same result with some invoices.  

I would appreciate any insights or suggestions on how to improve the accuracy of the AI model for invoice processing.

the is an example of the data recognition issue.
the  rows in red are incorrect 

 

Description Quantity Unit price Amount

|AGUACATE KG	2.100	5.50 |	11.55 |
|AJO TRENZA	10	2.00|	20.00 |
| ALBAHACA BANDEJAS	10	1.25 |	12.50 |
| APIO KILO	10	2.901	29.00 |
|CEBOLLINA MAZO	1	3.50|	3.50 |
LECHUGA GREEN LEAF KILO	6|	6.50|	39.00|
ROMANA KILO	6|LECHUGA	6.00|	36.00|
|LIMON AMARILLO UND	10	0.75|	7.50 |
|NARANJA IMP. UNIDAD	5	0.91|	4.55 |
|PAPA KG	25.200	3.15|	79.38|
|REPOLLO VERDE KILO	6.700	3.05|	20.44|
PAQ.	2|RABANO	3.75|	7.50 |
| TOMATE 3 X 3 KILO	25	4.70|	117.50 |
| TOMATE CHERRY BANDEJA	10	2.00	20.00 |
|ZANAHORIA KILO	10	2.35|	23.50|

Thank you!

[Antrod]

Hi @YanysM ,

Thanks for bringing that to our attention.

This seems to be caused by a misdetection of the words by the OCR service. It may be improved in the future as Microsoft is regularly improving its OCR capability but it can't be improved right now by AI Builder model training.

If you are running a Power Automate flow to extract the information, I would recommend to include a cleaning step after data extraction to handle such case. For example, if you detect "|" or even letters in the quantity field, you could assume that a part of the string is not in the correct place. You could therefore add a logic that turns 

ROMANA KILO

6|LECHUGA

to

LECHUGA ROMANA KILO

6

Hope that could help!