Dual-Encoder Neural Segmentation for Complex Document Indexing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for processing complex text documents, such as legal contracts and scientific articles, rely heavily on optical character recognition (OCR) followed by semantic analysis, but these methods often result in inaccurate indexing due to imperfect OCR, lacking contextual and positional information, and are inefficient in processing large volumes of documents.
Innovation Solution
A neural network architecture comprising two parallel encoding chains with residual connections and a decoding chain is used to process both an initial image and a mask image of a document, enabling precise semantic segmentation and improved indexing by identifying and classifying document zones like logos, headings, and signatures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If conventional OCR followed by semantic analysis is used, then document processing can be automated, but the indexing accuracy deteriorates due to imperfect OCR and lack of contextual information
Solution Approach 1:
The patent segments the document processing task into two parallel encoding chains: one processing the original image and another processing the OCR result image. This segmentation allows each chain to focus on specific aspects (visual structure vs. text content) and combine their strengths to achieve both automation and accuracy
Solution Approach 2:
The patent introduces an intermediary OCR result image as a bridge between the original image and the final semantic analysis. This intermediary contains both the text content and positional information, serving as a mediator that preserves contextual information while enabling automated processing
2Extent of automation
If conventional OCR followed by semantic analysis is used, then document processing can be automated, but the loss of contextual and positional information worsens the quality of semantic analysis
Solution Approach 1:
The patent adds a spatial dimension back into the text processing by creating an image from OCR results that preserves positional information. This dimensional transformation allows the system to maintain both automated text processing and spatial contextual information simultaneously
3Measurement precision
If conventional semantic analysis software is used, then document indexing can be performed, but the processing time increases and productivity decreases
Solution Approach 1:
The patent segments the processing into parallel independent encoding chains that can be computed simultaneously, reducing the overall processing time while maintaining high-quality semantic analysis through the combination of visual and text-based features
Data Source
AI summary
The invention relates to a neural network for semantic segmentation of a document with complex text. The network comprises a first multilayer neural encoding chain of an initial image file of the document to be processed, a second multilayer neural encoding chain of a mask image file of the document to be processed, a multilayer neural decoding chain connected to the outputs of the first and second encoding chains, a first bridge of parallel residual connections between the layers of the first encoding chain and the layers of the decoding chain, a second bridge of parallel residual connections between the layers of the second encoding chain and the layers of the decoding chain, the residual connections generating files of the same size that are connected after a layer of the neural decoding chain generating a file of the same size.