Machine Learning Document Classification via Multi-Classifier Agreement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for classifying scanned or captured images of physical paper documents face challenges due to variations in documents, particularly similar pages within multi-page documents, and incomplete metadata, leading to inefficiencies and inaccuracies in document classification using optical character recognition and keyword extraction.
Innovation Solution
A machine learning-based document classification method employing multiple classifiers in an iterative process, where a predetermined number of classifiers agree on a label, reducing the need to scan entire documents and improving accuracy by utilizing techniques like convolutional neural networks, term frequency-inverse document frequency, and gradient boosting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If optical character recognition and keyword extraction are used for document classification, then text can be extracted and categorized, but the system becomes slow and inefficient requiring extensive processor and memory resources
Solution Approach 1:
The patent divides the document into multiple image sections (e.g., header, body, footer regions) and classifies each section independently using machine learning models. This segmentation allows parallel processing of different regions, improving processing speed while maintaining overall classification accuracy through aggregation of section-level predictions.
2Productivity
If keyword extraction and natural language processing are used to classify documents, then text-based categorization can be performed, but the system becomes inaccurate when similar keywords appear in unrelated documents
Solution Approach 1:
The patent introduces machine learning models trained on document image features as an intermediary between raw document images and final classifications. These models learn to recognize document types based on visual patterns and structural features rather than relying solely on keyword presence, thereby reducing false positives from similar keywords in unrelated documents.
3Measurement precision
If the entire document is scanned for classification to ensure accuracy, then comprehensive analysis is achieved, but processor and memory utilization increases significantly
Solution Approach 1:
The patent extracts and processes only the most informative image sections of documents (such as title areas, header regions, or distinctive visual elements) rather than analyzing the entire document. This selective extraction reduces the amount of data requiring processing and memory storage while maintaining classification accuracy by focusing on discriminative features.
Data Source
AI summary
In some aspects, the disclosure is directed to methods and systems for machine learning-based document classification using multiple classifiers. Various classifiers may be employed during different iterations of the method to advance the classification of a document. The document may be classified and labeled in response to a predetermined number of classifiers agreeing upon a meaningful label. Further, the meaningful label may only be applied to the document in the event that the classifiers predicted the document label with a confidence score in excess of a threshold value.