Amazon Textract AnalyzeExpense vs docTR

Amazon Textract AnalyzeExpense extracts financial text and data from invoices and receipts through an API or web service. Results can include grouped line items and summary fields such as purchased items, prices, receipt headers, and vendor names. The API can normalize field names and column headers to help downstream applications compare documents. It can also identify vendors from logos and extract items, quantities, and prices when those details have no labels or column headers. Supply documents as image bytes or an Amazon S3 object; AWS advises PDF, JPG, or PNG files, and invoice and receipt processing is English only. AnalyzeExpense processes synchronously. Asynchronous invoice and receipt analysis uses StartExpenseAnalysis and GetExpenseAnalysis. The listed rate is $0.01 per page for the first 1 million pages per month in US West (Oregon), then $0.008 per page; custom high-volume pricing is available. New AWS customers can use the Free Tier for three months, up to 100 pages per month. AWS says inputs may be used to improve Textract and related AI technologies, with an Organizations opt-out policy.

docTR is a free deep-learning OCR library for finding and recognizing text in documents, aimed at document automation and research. Its pretrained OCR predictors first detect text and then recognize it, while a layout predictor identifies regions such as tables, figures, and headers. It reads PDFs and images; web-page input requires the optional HTML extra. Results can be rendered as plain text or exported as a JSON-serializable nested dictionary. The project describes its predictors as optimized for CPU and GPU inference. Users can train custom detection, recognition, layout, and table-structure models when pretrained models are not suitable. Documentation covers loading and sharing models through the Hugging Face Hub, and includes a minimal REST API deployment template and a browser demo. Installation is available through pip, Git, or official Docker images. The library requires Python 3.11 or higher, and pretrained weights download on first use and are cached locally. For AWS Lambda, the guide advises disabling multiprocessing and placing the model cache within /tmp. The project uses the Apache License 2.0.