---
title: "From Pixels to Text: Evaluating Open-Source OCR Models on Japanese Medical Documents - Bing Wang"
tags: PyConTW2025, 2025-organize, 2025-共筆
---

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

{%hackmd L_RLmFdeSD--CldirtUhCw %}

<iframe src=https://app.sli.do/event/hN7HWrRL23A3UpX8PnJhjF height=450 width=100%></iframe>

:::success
本演講提供 AI 翻譯字幕及摘要，請點選這裡前往 >> [PyCon Taiwan AI Notebook](https://pycontw.connyaku.app/?room=k2zPl6OrL8RGzZEVpoFS)
AI translation subtitles and summaries are available for this talk. Click here to access >> [PyCon Taiwan AI Notebook](https://pycontw.connyaku.app/?room=k2zPl6OrL8RGzZEVpoFS)
:::

> Collaborative writing start from below
> 從這裡開始共筆

## The Engines

### Paddel

- 文字歪斜的情況下成果不好
- 邊緣的文字容易出錯
- 多語言的情況下總是mix up with簡體中文

優點： 迅速, apache2.0 license, Versatile piplines

### Yomitoku

- 手寫或字體較特別時 容易出錯
- logo易出錯
- license 是 c.c. 不給商業用途

優點： Shows text detection confidence rate

在各項評估中 碾壓另兩個opensource ocr engine

### Tesseract

- 有複雜版面設計的文字時 效果不好
- No GPU support 費時較常是最大缺點

優點:

## 驗證方法

CER = S + D + I / N

CER(字元錯誤率)越低 和原始文字越接近

## How the texts are built to compare with ground truth

文字的串接順序 以辨識出的外框中線20pixel範圍內的就接起來

## 一些能讓效果提升的處理

- sharpened = cv2.filter2D(img, -1, kernel)
- contrast = cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)

經過上述處理的Image4 都有 Improved Error Rate

Below is the part that speaker updated the talk/tutorial after speech
講者於演講後有更新或勘誤投影片的部份