OCR reading issue with multiple pages

[rashidwiizb]

When I attempt to call the "/ocr/" API with a PDF containing 30, 50, or 80 pages, it takes more than 15-30 minutes to receive a response. Is there a way to optimize this process for faster response times?

The detection architecture (det_arch) used is 'fast_base', and the recognition architecture (reco_arch) is 'crnn_vgg16_bn', with default query parameters.

[felixdittrich92]

Hi @rashidwiizb 👋,

The API code we provide is only a reference ..so nothing you should use for any production use case :)

At the end the most speed up would be possible by using a Nvidia GPU

There are some other options but nothing which would save minutes in this case.

Otherwise the lazy way: split the pdfs before and merge the results

[felixdittrich92]

If you are limited to CPU check out: https://github.com/felixdittrich92/OnnxTR (it's much faster especially on CPU and requires less ressources)

[rashidwiizb]

So, how can I use this in production?