Skip to content

README.md

Latest commit

 

History

History
115 lines (83 loc) · 3.93 KB

README.md

File metadata and controls

115 lines (83 loc) · 3.93 KB

Paddle OCR Application

⚠️ Beta: This application is currently in beta. Features and APIs may change.

Run the OCR example:

hailo-ocr

To close the application, press Ctrl+C.

This example demonstrates Optical Character Recognition (OCR) using PaddleOCR models with Hailo AI acceleration. The pipeline detects text regions in video frames and recognizes the text content within those regions.

Key Features:

  • Text Detection: Identifies text regions in video frames using a detection model
  • Text Recognition: Recognizes and extracts text from detected regions
  • Real-time Processing: Processes live video streams with tracking support
  • Multi-stage Pipeline: Detection → Tracking → Cropping → Recognition

Running with Raspberry Pi Camera input:

hailo-ocr --input rpi

Running with USB camera input (webcam):

There are 2 ways:

Specify the argument --input to usb:

hailo-ocr --input usb

This will automatically detect the available USB camera (if multiple are connected, it will use the first detected).

Second way:

Detect the available camera using this script:

get-usb-camera

Run example using USB camera input - Use the device found by the previous script:

hailo-ocr --input /dev/video<X>

For additional options, execute:

hailo-ocr --help

Running as Python script

For examples:

python paddle_ocr.py --input usb

App logic

The OCR pipeline operates in two stages:

  1. Text Detection Stage: The detection model (PaddleOCR DB) identifies text regions in the frame, outputting bounding boxes around text areas.

  2. Text Recognition Stage: Each detected text region is cropped and fed to the recognition model (PaddleOCR CRNN), which outputs the recognized text string.

The pipeline includes a tracker to maintain consistent text region IDs across frames, reducing redundant recognition operations.

Working in Python with the results

The basic idea is to utilize the pipeline's callback function. In simple terms, it can be thought of as a Python function that is invoked at the end of the pipeline when frame processing is complete.

This is the recommended location to implement your logic.

def app_callback(element, buffer, user_data):
    roi = hailo.get_roi_from_buffer(buffer)
    text_detections = roi.get_objects_typed(hailo.HAILO_DETECTION)
    
    for detection in text_detections:
        if detection.get_label() == "text_region":
            # Get OCR text result from classification
            ocr_objects = detection.get_objects_typed(hailo.HAILO_CLASSIFICATION)
            if len(ocr_objects) > 0:
                text_result = ocr_objects[0].get_label()
                confidence = detection.get_confidence()
                print(f"Detected text: {text_result} (confidence: {confidence:.2f})")
    return

The user_app_callback_class extends the base callback class with OCR-specific functionality:

  • add_ocr_result(): Store detected text with confidence and bounding box
  • get_ocr_results(): Retrieve all detected text in the current frame
  • clear_ocr_results(): Clear results between frames

Pipeline Architecture

Source → OCR Detection → Tracker → Cropper → OCR Recognition → Callback → Display
  • OCR Detection: Identifies text regions using DB (Differentiable Binarization) model
  • Tracker: Maintains consistent IDs for text regions across frames
  • Cropper: Extracts detected text regions for recognition
  • OCR Recognition: Uses CRNN model to recognize text in cropped regions

Configuration

The OCR pipeline uses a configuration file located at local_resources/ocr_config.json which includes:

  • Frequency dictionary for spell-checking
  • Recognition parameters
  • Post-processing settings

All pipeline commands support these common arguments:

Common arguments