Pixel·OCR — From Scratch

A complete OCR web app built with FastAPI + pure Python OCR engine.
No OpenCV. No NumPy. No image processing libraries.

Project Structure

ocr_app/
├── main.py            ← FastAPI server + /api/ocr endpoint
├── ocr_engine.py      ← Full OCR pipeline (pure Python)
├── requirements.txt
└── templates/
    └── index.html     ← HTML/CSS/JS frontend

Setup & Run

# 1. Install dependencies (standard library only for OCR, FastAPI for server)
pip install -r requirements.txt

# 2. Start the server
uvicorn main:app --reload --port 8000

# 3. Open in browser
open http://localhost:8000

API

POST /api/ocr

Upload an image and get extracted text back.

Form fields:

Field	Type	Default	Description
file	file	—	PNG, BMP, or JPG image
threshold	int	auto	Binarization threshold (0–255)
row_gap	int	3	Max pixel gap between text rows
col_gap	int	5	Max pixel gap between characters

Response:

{
  "success": true,
  "text": "Hello World",
  "lines": 1,
  "words": 2,
  "characters": 10,
  "width": 640,
  "height": 480,
  "threshold_used": 128,
  "threshold_auto": 128,
  "elapsed_ms": 312.4
}

OCR Pipeline

1. Raw byte decoding — PNG chunk parsing with zlib inflate, BMP header parsing
2. Grayscale — ITU-R BT.601: 0.299·R + 0.587·G + 0.114·B
3. Otsu binarization — maximize inter-class pixel variance
4. Row segmentation — find horizontal ink-filled bands
5. Column segmentation — isolate individual character bounding boxes
6. Feature extraction — 26-feature vector per glyph
7. Template matching — weighted Euclidean distance to 80+ reference bitmaps

Tips for Best Results

• Use dark text on white/light background
• Works best on printed or typed text (not handwriting)
• Try adjusting the threshold slider if detection is poor
• PNG format gives the cleanest results