Forum Discussion

Copper Contributor

Sep 15, 2022

Form Recognizer response times averaging up to 9 secs.

I am using Form Recognizer, Custom models, single template to read handwriting from paper forms. I have chosen S0 tier and the best response time I am getting using the REST API call is around 5 secs...

tushargw

Copper Contributor

Sep 16, 2022

Virile , thanks a lot for writing an elaborate response. I will try to address every observation you made.

- Yes, the handwritten forms use a printed template and are sent as scanned images. So, @I, will experiment if reducing the DPI reduces the response time.

- I was leaving out handwriting quality and accuracy right now from the discussion to focus on performance. But @I will try to see if there is any correlation by visual inspection.

- The duration mentioned is when the call is made from Azure function to Form Recognizer Service in the same location within Azure and doesn't include the time to get the image to the FR service. Docker is an option I can experiment with, but currently, I am not compelled to prioritize it over other action items, but it will be in my list.

Thanks again for responding.

Virile

Brass Contributor

Sep 19, 2022

tushargw, I don't think that Docker nor changing the DPI will help much in this instance. I say this because of your response stating the action you're performing is an Azure function to the form recognizer service itself. I wouldn't expect a large amount of latency in this instance, however; I do suppose a lower DPI could help a little. Hope to see you get your problem figured out. Perhaps customer service?

tushargw
Copper Contributor
Sep 22, 2022
Virile, I have already raise a ticket with MS Support and it is in progress. The following changes didn't change the response time
- Changing image size
- Reducing fields to be extracted (business need will not permit this anyways)
- Removing extra text from the form (to reduce recognition)
- I did not change the DPI but I have images of different DPIs and that didn't show any correlation.
A new factor that surfaced in my research is image color depth. From what I have read is that black and white images (not even grey shades) work the best. Though it is not possible to change every image to black and white, I intend to try reducing the color depth to 8-bit from 24-bit and check.
Also I tested against prebuilt-layout model and still got similar response time. So looks like it is not even model dependent.