Automatic Product Creator for POS/ERP

Turn raw product photos into structured catalog data and e‑commerce‑ready hero images with a single Python script. The pipeline detects product regions, reads packaging text, masks backgrounds, and exports JSON plus white‑background hero crops.

Features

• YOLO‑World open‑vocabulary detection (products, labels, barcodes, etc.)
• EasyOCR line extraction with red overlay previews
• Masked “text‑only” preview to focus on readable regions
• Hero crops and cutouts on pure white backgrounds
• Barcode reads (ZXing‑CPP / pyzbar fallback)
• Combined JSON per product with summaries, dates, MRP, barcodes, ingredients, claims, and nutrition facts
• Optional LLM assist (OpenAI) to structure/enhance outputs
• Optional product API upsert via PRODUCT_API_URL

Quick start

1. Requirements

   • Python 3.11+
   • System zbar (for barcodes): brew install zbar (macOS) or your distro’s package
   • GPU optional; CPU works (slower)

2. Install deps

pip install -r requirements.txt

3. Add images
   Place product photos under input/products/<product_name>/ (one folder per product, up to 6 images). Example: input/products/my_product/image1.jpg.

4. Run

python ocr_pipeline.py

Artifacts will appear in output/.

Outputs

For each image you’ll see:

• <stem>_detect.jpeg — YOLO boxes
• <stem>_ocr.jpeg — OCR boxes/confidence
• <stem>_masked.jpeg — non‑text whited out (when OCR lines exist)
• <stem>_hero.jpeg — square crop on white
• <stem>_hero_cutout.jpeg — background‑removed square (if OpenCV available)
• <stem>.json — detections, OCR, barcodes, timings, output paths

For each product folder you’ll also get a combined JSON plus output/catalog.json and output/catalog.html for browsing results.

Example combined JSON snippet

{
  "summary": {
    "product_name": "Red Label Tea",
    "brand": "Brooke Bond",
    "net_weight": "450 g",
    "serving_size": "tsp (2g)",
    "expiry": "2025-06-15"
  },
  "barcode": {
    "items": [
      { "type": "EAN13", "data": "8720608614813" }
    ]
  },
  "mrp": { "amount": "299", "currency": "₹", "raw": "MRP ₹299.00 (incl. of all taxes)" },
  "dates": {
    "best_by": "2025-06-15",
    "raw_candidates": ["MFD 15/06/2024", "EXP 15/06/2025", "LOT A1B2C3"]
  },
  "views": {
    "front": { "hero": "output/product_front_hero.jpeg", "ocr": "output/product_front_ocr.jpeg" },
    "back":  { "hero": "output/product_back_hero.jpeg",  "ocr": "output/product_back_ocr.jpeg" }
  }
}

Configuration (env vars)

• MAX_SIDE (default 2000) — resize longest side before inference
• YOLO_MODEL (default yolov8m-world.pt)
• YOLO_CONF (default 0.25) / YOLO_IOU (default 0.45)
• YOLO_CLASSES — comma‑separated prompts
• OCR_LANGS (default en)
• HERO_SIZE (default 1024)
• CUT_PAD_PCT (default 0.06), CUT_ITERS (default 5)
• PRODUCT_API_URL — POST combined products to <url>/products
• OPENAI_API_KEY — enables LLM structuring/enhancement (leave empty to skip)

Repo hygiene

• output/, input/, node_modules/, caches, and editor files are git‑ignored to avoid noise and large assets.
• Sample only: samples/output_sample.json shows the combined output shape without shipping real data.
• Secrets: OPENAI_API_KEY is not hardcoded; set it in your environment if you use LLM features.

How it works (high level)

1. Downscale image to MAX_SIDE if needed.
2. Run YOLO‑World detections and EasyOCR lines.
3. Build overlays and masked preview; read barcodes.
4. Pick the best product box → hero and cutout crops.
5. Extract anchor info (dates/MRP) and heuristic summary.
6. (Optional) Send JSONs to OpenAI for structured output.
7. Write per‑image JSONs, combined product JSON, and catalog HTML.

Showcase

• Product catalog UI (catalog HTML with hero/JSON preview):
  
• AI email/resume automation dashboard (My-Email-MCP):
  

Notes

• If no product‑like box is found, hero images may be skipped (noted in JSON).
• If OpenCV is missing, the cutout hero is skipped (noted in JSON).
• OCR‑dependent outputs are omitted when no lines are detected.

Suggested GitHub topics

• computer-vision, ocr, yolo, easyocr, openai, fastapi, pgvector, commerce, product-catalog