Multi-modal Ensemble Deep Learning for Document Image Classification

The system enables documents to automatically classify and split themselves using deep learning models that autonomously analyze document images and determine page boundaries without human intervention. The multi-modal ensemble models process document features and generate predictions automatically, eliminating the need for manual feature engineering and expert knowledge input.

The patent replaces manual mechanical processes (human experts manually analyzing and splitting documents) with automated deep learning systems. Multiple independent trained models process document images through neural networks, substituting human cognitive work with computational algorithms that automatically identify document boundaries and classify pages.

The patent divides the document classification task into multiple independent models, each specializing in different aspects: image features, font features, margin features, and textual features. Each model processes specific modalities separately, then their predictions are combined through an additional trained model. This segmentation allows each component to focus on its strength while maintaining manageable complexity through modular architecture.

The system combines multiple independent model predictions into a composite classification system. Like composite materials that combine different substances to achieve superior properties, the ensemble of multiple models produces more accurate and reliable document classification than any single model could achieve alone, with the additional trained model synthesizing the combined predictions.