OCR Text Detection with Automatic Pipeline Selection

This repository implements a practical OCR (Optical Character Recognition) pipeline using Tesseract OCR + OpenCV.
Instead of using one fixed preprocessing setting, it automatically tries multiple preprocessing + PSM configurations and picks the best one per image using confidence-driven heuristics.

---

What this project produces

For each input image in:

• data/sample_images/

the system generates:

• Extracted text → outputs/text/<image_name>.txt
• Best preprocessing debug image → outputs/viz/<image_name>_<pipeline>_pre.png
• OCR bounding boxes visualization → outputs/viz/<image_name>_ocr_boxes.png

Note: The CSV file outputs/metrics/ocr_words.csv is also generated for analysis, but this README focuses on the visual results.

---

How to run

python run_ocr.py

---

Example run (your results)

Console output from your run:

Processed: captcha2.jpg | pipeline=adapt_deskew_psm11__UNRELIABLE | words_kept=1
Processed: numbers_gs150.jpg | pipeline=otsu_psm4 | words_kept=95
Processed: plaid_bw200.jpg | pipeline=otsu_psm4 | words_kept=20
Processed: stock_gs200q25.jpg | pipeline=adapt_psm11 | words_kept=200

---

Visual Results (with paths)

1) CAPTCHA-like image (flagged as unreliable)

Input

• data/sample_images/captcha2.jpg

Selected pipeline

• adapt_deskew_psm11__UNRELIABLE

Best preprocessed output
Path: outputs/viz/captcha2_adapt_deskew_psm11__UNRELIABLE_pre.png

OCR bounding boxes
Path: outputs/viz/captcha2_ocr_boxes.png

---

2) Handwritten + printed table

Input

• data/sample_images/numbers_gs150.jpg

Selected pipeline

• otsu_psm4

Best preprocessed output
Path: outputs/viz/numbers_gs150_otsu_psm4_pre.png

OCR bounding boxes
Path: outputs/viz/numbers_gs150_ocr_boxes.png

---

3) Poster / stylized text

Input

• data/sample_images/plaid_bw200.jpg

Selected pipeline

• otsu_psm4

Best preprocessed output
Path: outputs/viz/plaid_bw200_otsu_psm4_pre.png

OCR bounding boxes
Path: outputs/viz/plaid_bw200_ocr_boxes.png

---

4) Newspaper stock listings

Input

• data/sample_images/stock_gs200q25.jpg

Selected pipeline

• adapt_psm11

Best preprocessed output
Path: outputs/viz/stock_gs200q25_adapt_psm11_pre.png

OCR bounding boxes
Path: outputs/viz/stock_gs200q25_ocr_boxes.png

---

Notes / limitations (expected behavior)

• CAPTCHA-like images are designed to defeat OCR; the system flags these as __UNRELIABLE.
• Handwritten text is only partially supported with Tesseract; results vary by writing style and image quality.
• Decorative logos and highly stylized fonts may not be recognized perfectly.

---

Files to commit vs ignore (practical)

Recommended to commit:

• README.md (this file)
• src/, run_ocr.py, requirements.txt
• A small number of demo outputs under outputs/viz/ (like the ones shown above)

Recommended to ignore:

• Large/temporary outputs/ runs beyond a small demo subset
• Large datasets under data/ (keep only a few sample images)