Skip to main content

​
What is OCR?

OCR (Optical Character Recognition) converts images of text into machine-readable data. Instead of manually typing out what’s in a photo of a document, OCR does it automatically. Trace OCR is a REST API that does this for you. Upload an image or PDF, and get back structured text with positions and confidence scores — all in a single API call.

​
Use cases

Here are some concrete examples of what you can build with Trace OCR:
  • Receipt and invoice processing — extract line items, totals, dates, and vendor information from photos of receipts or scanned invoices
  • Document digitization — convert scanned papers, contracts, and reports into searchable, editable text
  • Form data extraction — pull field values from filled-in forms, applications, and questionnaires
  • ID and card reading — read information from ID cards, business cards, or membership cards
  • Compliance and archival — index large document archives so they become searchable and auditable

​
How it works

Trace OCR processes documents in three steps:
  1. Upload — you send an image or PDF to the API
  2. Detect — a deep learning model scans the image and finds all text regions
  3. Recognize — a second model reads each detected region and outputs the text
The result is a structured hierarchy: pages > blocks > lines > words. Each word comes with its bounding box coordinates and a confidence score telling you how certain the model is about the reading. For a deeper dive into the two-stage architecture, see OCR pipeline.

​
Example: processing a receipt

Let’s walk through a real example. Here’s a photo of a crumpled Apple Store receipt:
Photo of an Apple Store receipt
Send it to the API: 
curl -X POST https://ocr-api.trace.so/ocr/ \
  -F "[email protected]"
Trace OCR detects every text region in the image and draws bounding boxes around them:
Receipt with OCR bounding boxes overlaid
The API returns structured JSON with every word, its position, and confidence: 
[
  {
    "name": "receipt.png",
    "orientation": { "value": 0, "confidence": null },
    "language": { "value": null, "confidence": null },
    "dimensions": [1536, 1024],
    "pages": [
      {
        "blocks": [
          {
            "geometry": [0.214, 0.149, 0.777, 0.783],
            "detection_score": 0.55,
            "lines": [
              {
                "geometry": [0.386, 0.149, 0.641, 0.181],
                "words": [
                  {
                    "value": "RECEIPT",
                    "confidence": 0.99,
                    "geometry": [0.386, 0.149, 0.641, 0.181],
                    "detection_score": 0.57
                  }
                ]
              },
              {
                "geometry": [0.254, 0.208, 0.768, 0.230],
                "words": [
                  { "value": "APPLE", "confidence": 0.99, "geometry": [0.254, 0.208, 0.356, 0.227] },
                  { "value": "STORE,", "confidence": 0.99, "geometry": [0.365, 0.209, 0.478, 0.230] },
                  { "value": "GRAFTON", "confidence": 0.99, "geometry": [0.488, 0.211, 0.639, 0.229] },
                  { "value": "STREET", "confidence": 0.99, "geometry": [0.649, 0.212, 0.768, 0.229] }
                ]
              },
              {
                "geometry": [0.222, 0.485, 0.471, 0.502],
                "words": [
                  { "value": "MacBook", "confidence": 0.99, "geometry": [0.222, 0.485, 0.332, 0.500] },
                  { "value": "Pro", "confidence": 0.99, "geometry": [0.336, 0.485, 0.378, 0.501] },
                  { "value": "14-inch", "confidence": 0.99, "geometry": [0.383, 0.484, 0.471, 0.502] }
                ]
              }
            ]
          }
        ]
      }
    ]
  }
]
The response above is abbreviated. The full response includes every word on the receipt — addresses, dates, prices, line items, and totals — all with bounding boxes and confidence scores.
Here’s what each field means:
FieldDescription
valueThe recognized text for this word
geometryNormalized bounding box [x_min, y_min, x_max, y_max] where values range from 0 to 1 relative to page dimensions
confidenceHow certain the recognition model is about the text (0 to 1)
detection_scoreHow certain the detection model is that this region contains text (0 to 1)
In this example, the model correctly read “RECEIPT”, “APPLE STORE, GRAFTON STREET”, “MacBook Pro 14-inch”, the subtotal “€2,799.00”, and the total “€3,442.77” — all with confidence scores above 0.99.

​
Next steps

Quickstart

Make your first OCR request in minutes.

OCR pipeline

Understand the two-stage detection + recognition architecture.

Available models

Choose the right detection and recognition models for your use case.

API reference

Full endpoint documentation with parameters and response schemas.