From Pixels to Text: Evaluating Open-Source OCR Models on Japanese Medical Documents

Abstract

Strict data privacy policies in industries like healthcare prevent using commercial OCR services, creating hurdles to automatic document processing. At my company, for example, scanned files cannot use our auto-depersonalization pipeline and must be redacted manually, a time-consuming task. This challenge prompted us to seek an open-source OCR solution to be hosted in-house. We identified 3 engines that outperformed others. Using Python’s rapid development capabilities, we built a benchmark to assess each engine’s accuracy, memory usage, and speed with high-fidelity clinical documents. In addition, we used Python bindings to integrate C libraries such as OpenCV, for advanced image preprocessing that greatly enhanced OCR accuracy. This talk will demonstrate the pipeline, share the benchmarking results and highlight the trade-offs of each engine. Attendees will gain practical insights into how Python can be used to evaluate and optimize OSS OCR models in privacy-sensitive environments.

Description