A Hybrid Pytesseract–Easy OCR Pipeline with Semaphore-Controlled Asynchronous Serving for on-Premise Document Data Extraction

Organisations that handle invoices, bank forms and insurance papers still convert a great deal of paper into text by hand, and the automated tools they can buy either send documents to a third-party cloud or struggle once a scan gets noisy. We describe a small on-premise service that runs two open-source recognisers, PyTesseract and EasyOCR, on the same pre-processed image and keeps the more confident reading of each text segment. The service is written with FastAPI, and an asyncio semaphore caps how many recognition jobs run at once so that a burst of uploads queues politely rather than exhausting the machine. On a test set of 20 documents drawn from four categories the combined pipeline reached an average character accuracy of 94–96%, ahead of either engine used alone, and returned results in under a second per image when three or four uploads arrived together. Pre-processing contributed most on noisy scans. Nothing leaves the host, and there is no per-page fee. We report the design, the reasoning behind each algorithmic choice, the measurements, the limits of our evaluation, and how the numbers sit against three recent studies of OCR deployment.