Single-Stage Key-Value Extraction from Documents

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing OCR-based systems for extracting key-value pairs from documents and images are prone to errors, require additional processing and storage, and often ignore important context information due to independent training of models.

Innovation Solution

A single-stage, end-to-end system using machine learning techniques, specifically an object detection convolutional neural network algorithm, is employed to detect and localize data fields in images, eliminating the need for intermediate results and reducing the complexity of model training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a two-stage OCR system is used to extract key-value pairs, then character recognition and data classification can be performed separately, but errors in the first stage are propagated to the second stage and additional processing time is required

Engineering Contradiction:
Improvedata extraction accuracyVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent combines the character recognition stage and data classification stage into a single unified model. Instead of using separate OCR and classification models that operate sequentially, the invention trains a single model to directly output key-value pairs from document images, eliminating the propagation of errors between stages and reducing overall processing time.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent segments the document processing task into detecting key fields and extracting values, but integrates these operations into a single model rather than using separate sequential models. This allows the system to handle complex relationships between different data elements simultaneously while maintaining structured output.

Inventive Principle:
Principle #1Segmentation

2Ease of manufacture

If a two-stage OCR system is used, then intermediate results can be generated and processed, but additional file storage and computing resources are required

Engineering Contradiction:
Improvesystem implementabilityVSAvoidcomputing platform burden
Core Design Contradiction:
Ease of manufactureVSDevice complexity

Solution Approach 1:

The patent merges multiple processing stages into a single model, eliminating the need for intermediate result files and reducing computing resource requirements. The unified model processes the entire document image in one operation, producing final key-value pairs without requiring temporary storage or multi-stage computation.

Inventive Principle:
Principle #5Merging (Combining)

3Adaptability or versatility

If multiple models are trained independently for different document types, then each model can be optimized for its specific task, but context information from other models is lost

Engineering Contradiction:
Improvemodel specializationVSAvoidcontext information
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent trains a single universal model that can handle multiple document types and extraction tasks. This model incorporates contextual information from various document types during training, allowing it to maintain specialized knowledge for different document categories while preserving contextual relationships between different data elements across the entire document.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4567744A1Generative ai system and method for key and value pair information extraction from documents
Publication Date: 2025.06.11 SOCURE INC
  • EP4567744A1 patent drawingFigure 1
  • EP4567744A1 patent drawingFigure 2
  • EP4567744A1 patent drawingFigure 3

AI summary

A single stage, end to end system and methodology is employed to obtain key-value pairs from original documents and images without the requirement for intermediate results or data as would typically be required in OCR based data capture solutions. The system and methodology of the present invention requires only a single model to generate key-value pairs from original documents and images as opposed to the use of at least two models which are required when traditional OCR data capture solutions are implemented.