Image Processing¶
Featrix can understand images referenced by URL in your datasets, turning product photos, document scans, user avatars, and any other images into rich features that interact with all your other columns.
The Problem¶
Many real-world datasets contain image URLs — product photos in e-commerce catalogs, property photos in real estate listings, document scans in insurance claims, profile pictures in user databases. Traditional tabular ML ignores these entirely or requires a separate computer vision pipeline with manual feature engineering.
How Featrix Solves It¶
When image understanding is enabled, Featrix automatically:
- Detects which columns contain image URLs
- Downloads and analyzes each image using three complementary models
- Learns how image content relates to every other column in your data
No separate vision pipeline. No manual feature engineering. Images become first-class features alongside your text, numbers, and categories.
What Gets Extracted¶
| Model | What It Captures | Example |
|---|---|---|
| DINOv2 (visual) | Overall visual content — composition, color, texture, style | Two similar product photos have similar embeddings |
| EasyOCR (text) | Any text visible in the image — labels, prices, signs | "Nike Air Max $129.99" read from a shoe label |
| ResNet (objects) | What objects are in the image, with confidence scores | {"running_shoe": 0.92, "sneaker": 0.05} |
Cross-Column Relationships¶
The real value is in how Featrix connects image content to your other data:
-
Image + Text: OCR text from images is matched against text columns using word overlap, substring matching, and semantic similarity. A brand name on a product label matches the
brandcolumn. -
Image + Categories: Detected objects are semantically matched against category values. ResNet detecting "laptop" correlates with a
categorycolumn containing "Electronics". -
Image + Numbers: Numbers found via OCR (prices, quantities) are compared to numeric columns. A receipt showing "$42.99" matches a
total_amountof 42.99. -
Image + Image: When multiple image columns exist, visual similarity, shared text, and overlapping objects create natural relationships.
Example Scenarios¶
E-Commerce Product Catalog¶
| Column | Type | Example |
|---|---|---|
product_image |
IMAGE_URL | https://cdn.shop.com/shoes/nike-air-max.jpg |
brand |
Set | Nike |
category |
Set | Running Shoes |
price |
Scalar | 129.99 |
description |
String | "Nike Air Max 90 running shoe, white/black" |
Featrix discovers: the shoe label says "Nike" (matches brand), the image looks like athletic footwear (matches category), "$129.99" is visible on the tag (matches price), and the description overlaps with OCR text.
Real Estate Listings¶
| Column | Type | Example |
|---|---|---|
photo_url |
IMAGE_URL | https://photos.realty.com/house-123.jpg |
property_type |
Set | House |
price |
Scalar | 450000 |
sqft |
Scalar | 2200 |
Featrix learns: houses with pools, large yards, and modern kitchens (visual features) correlate with higher prices. Property type maps to distinctive visual patterns.
Document Processing¶
| Column | Type | Example |
|---|---|---|
scan_url |
IMAGE_URL | https://docs.company.com/invoice-456.png |
doc_type |
Set | Invoice |
amount |
Scalar | 1250.00 |
vendor |
String | "Acme Corp" |
Featrix reads: OCR extracts "Invoice", "$1,250.00", and "Acme Corp" from the scan, matching all three columns directly.
Getting Started¶
Add to your config.json:
Featrix handles the rest. For full configuration options and technical details, see Image Understanding.