LightOCRX

LightOCRX is a lightweight C++ inference engine for PaddleOCR ONNX models (v4 and v5). It prioritizes speed, low memory footprint, and portability, making it ideal for low-end devices.

---

This program depends on:

• libharu (PDF library)
• onnxruntime (ONNX model inference runtime)
• C++17 compiler

Please install both libraries before compiling.

Usage

lightocrx <detection_model.onnx> <recognition_model.onnx> <images...> [options]

Arguments

• detection_model.onnx
  Path to detection model

• recognition_model.onnx
  Path to recognition model

• images...
  One or more PNG/JPG image files or a directory

---

Options

• -boxt
  Box confidence threshold (default: 0.6)

• -bint
  Binarization threshold (default: 0.3)

• -unclip
  Unclip ratio (default: 2.0)

• -targetsize <int>
  Target size for detection (default: 640).
  Higher values improve accuracy at the cost of performance.

• -savecrops <0/1>
  Save cropped text regions (default: 0).
  Creates a folder named cropped containing cropped PNG images.

• -debugboxes <0/1>
  Draw detected boxes on output image (default: 0)

• -debugout <path>
  Output path for debug image (default: detected_boxes.png)

• -dict <path>
  Path to dictionary file (default: dict.txt)

• -targetw <int>
  Target width for recognition (default: 0)
  0 = dynamic
  -1 = use crop aspect ratio

• -targeth <int>
  Target height for recognition (default: 48)

• -pdf <path>
  Export as searchable PDF (optional).
  Creates a multi-page PDF when multiple images are provided.

• -minconf <float>
  Minimum confidence for PDF export (default: 0.3)

• -pdftext
  Export PDF as text-only (white background, black visible text, no images)

• -text <path>
  Export recognized text to a plain text file

• -textcls <0/1>
  Enable/disable text orientation classification and 180° rotation correction (default: 0).
  Useful when images contain rotated text.

• -clsmodel <path>
  Path to rotation classification model (required if -textcls 1)

• -h, --help
  Show help message

Examples

Basic usage:

lightocrx det.onnx rec.onnx image1.jpg image2.jpg -boxt 0.2 -dict mydict.txt

Save crops and debug boxes:

lightocrx det.onnx rec.onnx *.jpg -savecrops 1 -debugboxes 1

Process all images in a directory and export PDF:

lightocrx det.onnx rec.onnx images/ -pdf output.pdf

Export text-only PDF:

lightocrx det.onnx rec.onnx image1.jpg image2.jpg -pdf output.pdf -pdftext

Export recognized text to file:

lightocrx det.onnx rec.onnx image1.jpg image2.jpg -text output.txt

Enable rotation correction:

lightocrx det.onnx rec.onnx image.jpg -textcls 1 -clsmodel cls.onnx