Hugging Face's logo

Spaces:
chelleboyer
/
planolyzer-v2
Build error

App Files Community
planolyzer-v2 / ABOUT.md
chelleboyer's picture
chelleboyer
Initial commit: Planolyzer v2
1c47c79
preview code
|
raw
history blame contribute delete
2.63 kB

🛍️ Planolyzer Application Overview

Purpose

Planolyzer is an AI-powered tool designed to analyze retail shelf images and automatically check planogram compliance. It detects:

  • Empty spaces (missing products)
  • Wrong items (misplaced products)
  • Brand mismatches (using OCR and AI vision)

How It Works

1. Startup & Reference Data

  • On launch, Planolyzer loads:
    • A planogram JSON file describing the expected products, their positions, and the reference image filenames.
    • Reference product images for each SKU from a directory.
    • The reference planogram image (the “ideal” shelf).

2. User Interaction

  • The user is greeted with instructions and sample images.
  • The user uploads a photo of a retail shelf.

3. Image Validation

  • The app checks that the uploaded image is valid and readable.

4. Planogram Comparison

  • The uploaded image is compared to the reference planogram image using the CLIP model to ensure it’s a valid shelf photo.
  • If the similarity is too low, the user is asked to upload a better image.

5. Spot-by-Spot Analysis

For each product spot defined in the planogram:

  • The app crops the corresponding region from the uploaded image.
  • It checks if the spot is empty (using brightness and saturation thresholds).
  • If not empty, it:
    • Compares the crop to the expected reference image using CLIP (image-to-image similarity).
    • Compares the crop to all reference images to identify the most likely product if the item is wrong.
    • Uses CLIP text prompt matching (image-to-text) to further identify the product.
    • Runs OCR (EasyOCR) on the crop to extract any visible brand name or text.
    • Fuzzy matches the OCR result to known brands for robust detection.

6. Reporting

  • The app generates a detailed report for the user, including:
    • Which spots are empty (❌)
    • Which spots have the wrong item (⚠️), what the item most likely is, and what text was detected
    • A summary of empty and wrong spots as counts and percentages

7. User Feedback

  • The app provides real-time feedback with “Just a moment…” and “Analyzing…” messages while processing.
  • Results are displayed as soon as analysis is complete.

Key Technologies Used

  • CLIP (Contrastive Language-Image Pretraining): For image and text-based product recognition.
  • EasyOCR: For reading brand names and text from product images.
  • Chainlit: For the interactive chat/web interface.
  • OpenCV & PIL: For image processing and cropping.