Document Layer Separation for OCR and Non-Text Content Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document processing technologies face challenges in accurately extracting and handling non-textual content, such as logos, charts, and signatures, which can interfere with OCR accuracy and often fail to capture important information due to overlapping or mixed background colors.
Innovation Solution
A machine learning-based system separates document content into logical layers by content type, using a modified U-Net convolutional neural network, and applies specific handlers for each layer to process and extract relevant information, including text, logos, and barcodes, with context-sensitive cross-referencing for enhanced accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR is used to process documents with non-textual content, then text recognition can be performed, but OCR accuracy deteriorates due to overlapping elements and non-textual content interference
Solution Approach 1:
The patent segments the document image into multiple semantic layers (e.g., text layer, table layer, chart layer, watermark layer) based on content type and spatial relationships. This segmentation allows traditional OCR to process only the text layer while other layers are handled by specialized modules, thereby improving OCR accuracy by eliminating interference from non-textual content.
Solution Approach 2:
The patent introduces a multi-layer processing framework where an intermediary layer separation module divides the complex document into manageable semantic layers. This intermediary structure enables specialized handlers to process each layer appropriately, with the text layer being processed by OCR and other layers by dedicated content extraction modules, thus resolving the interference problem.
2Loss of information
If traditional single-layer document processing is used, then processing simplicity is maintained, but important non-textual information is lost
Solution Approach 1:
The patent segments the document into multiple semantic layers (text, table, chart, watermark) and assigns specialized handlers to each layer. This segmentation enables comprehensive extraction of both textual and non-textual information while maintaining modular processing architecture, thus capturing important non-textual information without excessive system complexity.
Solution Approach 2:
The patent creates a universal multi-layer processing framework where a single system can handle multiple content types through layer-specific handlers. The framework is designed to be extensible, allowing new content types to be added as separate layers without redesigning the entire system, thus achieving comprehensive information capture with manageable complexity.
3Adaptability or versatility
If color-based document segmentation is used, then content separation can be achieved, but the method fails on black and white documents or documents without sufficient color information
Solution Approach 1:
The patent changes the segmentation parameters from color-based to a combination of spatial position, content type classification, and visual characteristics. This parameter transformation enables the system to effectively segment documents regardless of color information, making it reliable for both colored and black-and-white documents, as well as documents with insufficient color data.
Solution Approach 2:
The patent creates a universal segmentation approach that works across diverse document types (colored, black-and-white, scanned, printed) by using multiple feature dimensions beyond color. This multi-functional segmentation mechanism ensures reliable content separation across all document types, significantly improving adaptability and reliability.
4Measurement precision
If simple background extraction is used, then processing speed is maintained, but accuracy deteriorates for documents with different or mixed background colors
Solution Approach 1:
The patent segments the background extraction task into multiple semantic layers, where each layer is processed independently based on its content type. This allows the system to maintain high processing speed by parallel processing while achieving high accuracy through specialized extraction methods for each layer, including intelligent handling of mixed and varying background colors.
Solution Approach 2:
The patent changes the background extraction approach from simple color-based thresholding to a multi-parameter system considering spatial relationships, content type classification, and visual characteristics. This parameter enhancement enables accurate extraction of backgrounds with different or mixed colors while maintaining processing efficiency through optimized algorithms for each semantic layer.
Data Source
AI summary
A method for processing electronic documents, comprising: receiving an electronic document; recognizing one or more content components in the electronic document; identifying a content type for each of the recognized content components; creating, by a layer separator, one or more logical layers from the recognized content components such that each of the logical layer contains only the content components of the same content type; and invoking a content-type specific content handler for each of the logical layers created. The layer separator comprises a machine learning (ML) model based on a modified U-Net convolutional neural network and trained to classify the content types of the content components. The modified U-Net CNN is improved over traditional U-Net CNN with transformers at each layer to achieve high recovery rate.