Skip to main content

​
What is text detection?

Text detection is the first stage of the OCR pipeline — it finds where text appears in an image without reading it. The /detection endpoint returns bounding box coordinates for every text region it finds. Use detection on its own when you only need to know where text is, not what it says:
  • Document layout analysis — identify text blocks, headers, and table cells by their positions
  • Text region highlighting — draw attention to text areas in a UI
  • Pre-processing for custom pipelines — feed detected regions into your own recognition model or downstream logic
  • Counting text blocks — quickly assess how much text is on a page

​
Example: detecting text in an invoice

Here’s an invoice PDF we’ll use to demonstrate detection:
Sample invoice with header, addresses, line items, and totals
Send it to the detection endpoint: 
curl -X POST https://ocr-api.trace.so/detection/ \
  -F "[email protected]"
The model scans the full page and draws bounding boxes around every text region:
Invoice with blue bounding boxes overlaid on every text region
The API returns a flat list of bounding boxes for each uploaded file: 
[
  {
    "name": "invoice.pdf",
    "bounding_boxes": [
      [0.0733, 0.0283, 0.3288, 0.0596],
      [0.8065, 0.0645, 0.9294, 0.0859],
      [0.5842, 0.1396, 0.6339, 0.1514],
      [0.6353, 0.1396, 0.6906, 0.1504],
      [0.7582, 0.1396, 0.8231, 0.1504],
      [0.0706, 0.1465, 0.1148, 0.1572],
      [0.0706, 0.1611, 0.1175, 0.1729],
      [0.1272, 0.1611, 0.1714, 0.1729],
      [0.1728, 0.1611, 0.2294, 0.1719],
      [0.5828, 0.1582, 0.6257, 0.1699]
    ]
  }
]
The response above is abbreviated — this invoice produces 111 bounding boxes in total. Each box corresponds to a single word-level text region.
Each bounding box is an array of four normalized coordinates: [x_min, y_min, x_max, y_max], where values range from 0 to 1 relative to the page dimensions. To convert to pixel coordinates, multiply by the page width and height. See Geometry format for details.

​
Parameters

All parameters are optional query parameters passed in the URL.
ParameterDefaultDescription
detection_modeldb_resnet50The detection architecture to use. See available models for options.
assume_straight_pagestrueReturn axis-aligned boxes. Set to false for rotated documents to get 4-point polygons instead.
preserve_aspect_ratiotruePad the image to preserve its aspect ratio before feeding it to the model.
symmetric_paddingtruePad symmetrically (centered) rather than bottom-right only.
detection_batch_size2Number of pages processed in parallel. Increase for multi-page PDFs if you have enough memory.
binary_threshold0.1Pixel-level threshold for the segmentation heatmap.
box_threshold0.1Minimum confidence to keep a detected box.

​
Tuning thresholds

The two threshold parameters control the sensitivity/precision trade-off:
  • Lower thresholds — detect more text regions, including faint or low-contrast text, but may introduce false positives
  • Higher thresholds — detect fewer, higher-confidence regions, reducing noise but potentially missing subtle text
# More aggressive detection (catch faint text)
curl -X POST "https://ocr-api.trace.so/detection/?binary_threshold=0.05&box_threshold=0.05" \
  -F "[email protected]"

# More conservative detection (reduce noise)
curl -X POST "https://ocr-api.trace.so/detection/?binary_threshold=0.3&box_threshold=0.3" \
  -F "[email protected]"

​
Detection vs full OCR

The /detection endpoint returns only bounding boxes — it tells you where text is but not what it says. The /ocr endpoint runs both detection and recognition, returning a full nested hierarchy of pages, blocks, lines, and words with recognized text and confidence scores.
/detection/ocr
OutputFlat list of bounding boxesNested hierarchy with recognized text
SpeedFaster (one model)Slower (two models)
Response sizeSmallerLarger
Use whenYou only need text locationsYou need to read the text
For a full walkthrough of the OCR pipeline, see OCR pipeline.

​
Next steps

Full OCR endpoint

Run detection and recognition together to get the full text content.

OCR pipeline

Understand the two-stage detection + recognition architecture.

Available models

Choose the right detection model for your use case.

API reference

Full endpoint documentation with parameters and response schema.