- 最新
- 投票最多
- 评论最多
The amazon-textract-transformer-pipeline sample shows a scalable batch PDF-to-images converter you might customize for this use case: Check out the code in the notebooks/preproc subfolder and the usage in the "Extract clean input images" section of notebook 1.
The current implementation is based on poppler, pdf2image and processes batches of documents (from Amazon S3) through SageMaker Processing. It's probably not the most efficient possible (Python...), but can scale up to bigger instances (via multiprocessing) and out to multiple instances (via data sharding).
If you needed (near)-real-time processing instead of batch, you could probably get a similar solution running on a containerized Lambda function (poppler requires a lower-level install than pip). In our draft upgrade branch, we instead use SageMaker Asynchronous Inference for this... Because request/response payload sizes and memory could be theoretically very large for documents with many pages.
