/easyocr_text_extraction

Primary LanguageJupyter Notebook

Extracting Numeric Values from Images using EasyOCR for Entity Recognition 🖼️🔢

Introduction 🚀

This project focuses on extracting numerical values and their corresponding units from a set of images, each associated with specific entities such as weight, height, or voltage. The goal is to develop a machine learning pipeline that accurately predicts these values and units from the images, overcoming challenges like variable image quality, different text fonts, and OCR (Optical Character Recognition) errors. 🧐

Approach 🛠️

Our approach involves several key steps:

  1. Image Text Extraction using OCR 📄:

    • We use EasyOCR with GPU acceleration to extract text from images efficiently.

  2. Text Pre-processing 🧹:

    • Clean and normalize the extracted text to reduce noise and correct common OCR misrecognitions.

  3. Value and Unit Extraction 🎯:

    • Using regular expressions tailored to each entity, we extract numerical values and units from the pre-processed text.

  4. Model Evaluation 📊:

    • Compare the extracted values and units with the ground truth to evaluate accuracy.

  5. Prediction on Test Data 🧠:

    • Apply the pipeline to the test dataset and generate predictions.

Models Used 🧩

  • EasyOCR: An open-source OCR tool optimized for GPU usage. It supports multiple languages and efficiently extracts text from images, which is crucial for our task.

Experiments and Iterations 🔄

Initial OCR Extraction ⚙️

  • EasyOCR was implemented to extract text from images.
  • Observed that raw OCR results often contained noise and misrecognized characters.

Text Pre-processing Enhancements 🧽

  • Lowercasing: Converted all text to lowercase for uniformity.
  • Character Replacement: Fixed common misrecognized characters (e.g., "O" → "0", "l" → "1").
  • Removing Unwanted Characters: Kept only alphanumeric characters, percentages, periods, and spaces.
  • Whitespace Normalization: Streamlined text by removing extra spaces.

Value and Unit Extraction Improvements 🔍

  • Used an entity-specific unit map to associate entities with their allowed units.
  • Created regex patterns to match numerical values followed by units.
  • Implemented fallback mechanisms to handle cases where initial extraction fails by assigning default units based on the entity.

Handling OCR Errors ⚠️

  • Introduced a clean_up_text function to handle OCR misrecognitions.
  • Adjusted regex patterns to account for values and units without spaces (e.g., "74m").

Model Evaluation 📐

  • Calculated training accuracy by comparing extracted values and units with ground truth.
  • Observed accuracy improvements with each iteration of text pre-processing and extraction enhancements.

Conclusion 🎉

Through systematic refinement of our OCR extraction and text pre-processing methods, we've significantly improved the accuracy of numerical value and unit extraction. The final pipeline adeptly handles common OCR errors and variations in text formats, providing a robust solution for entity recognition tasks. ✨