Text OCR Model

AI Data Capture SDK 4.0

Overview

Model Name: text-ocr-recognizer

Text OCR is used to detect and recognize text within a given image. It is capable of handling use cases involving printed text in various sizes and non-stylized fonts, as well as dot peen fonts and handwriting.​

Note: Zebra strongly recommends updating to the latest models.

Version History

New in v2.9.0

  • Improved accuracy across multiple use cases and data sets, particularly:
    • Longer Strings - If the updated model does not yield the required results, utilizing the tiling settings is recommended as an alternative.
    • Unexpected String Splitting - The new model mitigates unwanted splits caused by subtle character spacing differences. If further adjustments are necessary, utilizing the grouping settings is recommended.
    • Direct Part Marking (DPM) - Character accuracy is improved for select DPM use cases, primarily for single-word metal DPM images.
  • Changes to unclip ratio: Due to changes in the Text OCR model the unclip ratio range has shifted from 1.5-2.0 to 0.4-0.7 (0.6 is the default). Applications using non-default values should be updated to this new range. Use the following guidance for mapping values:
    • A previous value of 1.5 (original default) should be updated to 0.6.
    • A previous value of 2.0 value should be updated to 0.7.
    • Values between 1.5 and 2.0 should be scaled proportionally across the new 0.4 to 0.7 range.

New in v2.8.1

  • Improved overall performance, leading to increased accuracy.
  • Added support for Q-6690 device platform.

Requirements

  • Operating System: Android 14 or higher. For specific OS versions, refer to AI Data Capture SDK Release Notes from the Zebra support portal.

  • Minimum SDK Version: AI Data Capture SDK v4.0.0 or later.

  • Supported Zebra Devices:

    Features Platform Device Model
    Products with DSP
    Fastest and most battery efficient    		 QC6490 TC53, TC58, TC73, TC78, ET60, ET65
    QC5430 EM45
    Q-6690 ET401, TC501, TC701
    Products without DSP
    Applies to still images     		QC4490 TC53e, TC58e, MC9400, MC9450

    For more information on devices based on platform, see Zebra Platform Devices.

  • Memory Requirements: Since running multiple on-device models concurrently is highly resource-intensive, using multiple models within the AI Data Capture SDK is recommended for high-memory devices only.

Performance Details

Model Model Dimensions Typical Load Time (ms) Typical Inference Time (ms) Typical Maximum Read Distance (cm)
8pt font (Times New Roman)		 Typical Maximum Read Distance (cm)
12pt font (Times New Roman)		 Typical Maximum Read Distance (cm)
36pt font (Times New Roman)
text-ocr-recognizer​ 640x640 1580 110 No Read 20 55
text-ocr-recognizer​ 1280x1280 1780 180 25 40 130
text-ocr-recognizer​ 1600x1600 1740 270 35 40 145
text-ocr-recognizer​ 2560x2560 1840 480 50 65 165

Measurements Notes:

  • All measurements are estimations​.
  • Measurements were conducted using the Zebra TC53 device equipped with the Qualcomm 6490 chipset operating on the DSP AI accelerator​.
  • Images were 4MP Resolution under 300 Lux​ lighting conditions.
  • Read distance met at least 90% Recall.

Resources