Pipelines

MyOCR pipelines orchestrate multiple components (predictors, models) to perform end-to-end OCR tasks. They provide a high-level interface for processing images or documents.

Available Pipelines

CommonOCRPipeline

Defined in myocr/pipelines/common_ocr_pipeline.py.

This pipeline performs standard OCR: text detection, optional text direction classification, and text recognition.

Initialization:

from myocr.pipelines import CommonOCRPipeline
from myocr.modeling.model import Device

# Initialize pipeline for GPU (or 'cpu')
pipeline = CommonOCRPipeline(device=Device('cuda:0'))

Configuration:

The pipeline loads configuration from myocr/pipelines/config/common_ocr_pipeline.yaml. This file specifies the paths to the ONNX models used for detection, classification, and recognition relative to the MODEL_PATH defined in myocr.config.

# Example: myocr/pipelines/config/common_ocr_pipeline.yaml
model:
  detection: "dbnet++.onnx"
  cls_direction: "cls.onnx"
  recognition: "rec.onnx"

Processing:

The __call__ method takes the path to an image file.

image_path = 'path/to/your/image.png'
ocr_results = pipeline(image_path)

if ocr_results:
    # Access recognized text and bounding boxes
    print(ocr_results)

Workflow:

1. Loads the image.
2. Uses TextDetectionPredictor to find text regions.
3. Uses TextDirectionPredictor to classify the orientation of detected regions.
4. Uses TextRecognitionPredictor to recognize the text within each oriented region.
5. Returns a result object containing bounding boxes, text, and potentially confidence scores (details depend on the Predictor implementation).

StructuredOutputOCRPipeline

Defined in myocr/pipelines/structured_output_pipeline.py.

This pipeline extends CommonOCRPipeline by adding a step to extract structured information (e.g., JSON) from the recognized text using a large language model (LLM) via the OpenAiChatExtractor.

Initialization:

Requires a device and a Pydantic model defining the desired JSON output schema.

from myocr.pipelines import StructuredOutputOCRPipeline
from myocr.modeling.model import Device
from pydantic import BaseModel, Field

# Define your desired output structure
class InvoiceData(BaseModel):
    invoice_number: str = Field(description="The invoice number")
    total_amount: float = Field(description="The total amount due")
    due_date: str = Field(description="The payment due date")

# Initialize pipeline
pipeline = StructuredOutputOCRPipeline(device=Device('cuda:0'), json_schema=InvoiceData)

Configuration:

This pipeline loads its specific configuration from myocr/pipelines/config/structured_output_pipeline.yaml, which includes settings for the OpenAiChatExtractor (LLM model name, API base URL, API key).

# Example: myocr/pipelines/config/structured_output_pipeline.yaml
chat_bot:
  model: "gpt-4o"
  base_url: "https://api.openai.com/v1"
  api_key: "YOUR_API_KEY"

Processing:

The __call__ method takes an image path.

image_path = 'path/to/your/invoice.pdf'
structured_data = pipeline(image_path)

if structured_data:
    print(structured_data)

Workflow:

1. Performs standard OCR using the inherited CommonOCRPipeline to get the raw recognized text.
2. If text is found, it passes the text content to the OpenAiChatExtractor.
3. The extractor interacts with the configured LLM, providing the text and the desired json_schema (Pydantic model) as instructions.
4. The LLM attempts to extract the relevant information and format it according to the schema.
5. Returns an instance of the provided Pydantic model populated with the extracted data.

Performance Optimization

Batch Processing

# Process multiple images
results = [pipeline(img_path) for img_path in image_paths]

GPU Acceleration

# Use GPU for faster processing
pipeline = CommonOCRPipeline("cuda:0")

Memory Management

# Clear GPU memory
import torch
torch.cuda.empty_cache()

Error Handling

Pipelines handle various error cases:

• Invalid image format
• Missing model files
• GPU out of memory
• Invalid configuration

See the Troubleshooting Guide for common issues and solutions.

Comments