Pearson Correlation-Based Feature Selection for Document Classification Using Balanced Training.

Area of Science:

• Computer Science
• Information Science
• Machine Learning

Background:

• Digital document storage is prevalent, necessitating efficient retrieval methods.
• Traditional document handling is impractical, uneconomical, and ecologically unsound.
• Adverse image issues like signatures and handwritten notes hinder accurate digital document classification.

Purpose of the Study:

• To present a real-time supervised learning technique for document classification using deep convolutional neural networks (DCNNs).
• To mitigate the impact of image imperfections on document classification accuracy.
• To enhance the efficiency and reliability of retrieving digitally stored documents.

Main Methods:

• A novel data augmentation technique was developed, normalizing imbalanced datasets with the RVL-CDIP dataset.
• Deep convolutional neural network (DCNN) features were extracted using VGG19 and AlexNet models.
• Feature fusion and selection were performed using a Pearson correlation coefficient-based method to remove redundancy.

Main Results:

• The proposed technique achieved a classification accuracy of 93.1% on the Tobacco3482 dataset.
• The method effectively reduced the impact of adverse document image issues.
• Feature optimization significantly improved classification performance.

Conclusions:

• The developed DCNN-based technique offers a robust solution for digital document classification.
• The novel data augmentation and feature selection methods enhance classification accuracy.
• This approach validates the effectiveness of deep learning for practical document image analysis.