CNN and YOLO Financial Statement Extraction Across Document Formats

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for extracting data from financial documents, such as balance sheets, income statements, and cash flow statements, face challenges due to lack of standardization, low accuracy, and inefficiency in adapting to different document formats, requiring manual intervention and reconfiguration for each new type.

Innovation Solution

A machine learning-based computing system using convolutional neural networks (CNN) and You Only Look Once (YOLO) architecture to classify tabular and non-tabular data, combined with optical character recognition (OCR) for automated data extraction from financial documents, enabling adaptation to various formats and improving accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If generic document parsing software and OCR systems are used to process financial statements, then data extraction can be automated, but accuracy and processing speed remain low

Engineering Contradiction:
Improvedata extraction automationVSAvoiddata extraction accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent transforms the input document into multiple parameter variations including different image formats, resolutions, and preprocessing states. The system generates multiple versions of the same document with different parameters (e.g., grayscale vs. color, different DPI, cropped vs. full-page) to feed into the neural network, allowing the model to learn robust feature extraction across parameter variations and improve accuracy while maintaining automation.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If parsing systems are customized for specific clients or document formats, then processing accuracy improves, but adaptability to handle financial documents from different clients decreases

Engineering Contradiction:
Improveprocessing accuracyVSAvoiddocument format adaptability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent implements a universal deep learning-based parsing system that can handle multiple document formats and styles without customization. The neural network is trained on diverse financial document data during the learning phase, enabling it to universally process different document types (income statements, balance sheets, cash flow statements) from various clients. The system extracts structural information and financial data across different formats using the same core algorithm, achieving both accuracy and adaptability.

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

Solution Approach 2:

The patent performs preliminary learning and training phases where the system studies various document formats, structures, and patterns before actual data extraction. During this preliminary action phase, the neural network is trained on labeled financial documents to learn formatting variations, table structures, and data patterns. This pre-learning enables the system to adapt to different client documents without requiring customization, as the foundational knowledge is already embedded in the trained model.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If manual verification and entry by financial analysts is performed, then data accuracy improves, but processing time increases

Engineering Contradiction:
Improvedata extraction accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent implements a self-verification mechanism where the deep learning system automatically validates its own extractions by cross-checking financial relationships (e.g., verifying that assets equal liabilities plus equity in balance sheets, checking consistency across multiple statements). The system performs self-correction and validation without human intervention, maintaining high accuracy while eliminating the time-consuming manual verification step. The automated system serves itself by detecting and correcting errors through learned financial knowledge.

Inventive Principle:
Principle #25Self-service

4Measurement precision

If reconfiguration of parsing systems is performed for new financial statement types, then processing accuracy is maintained, but productivity decreases

Engineering Contradiction:
Improveprocessing accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent implements a dynamic system where the parsing configuration automatically adapts to new document types through the trained neural network's learned parameters. Instead of static, manually-configured parsers, the system dynamically adjusts its extraction rules, structural expectations, and validation criteria based on the input document characteristics. The deep learning model's weights and biases are dynamically applied to handle various statement types, maintaining accuracy without requiring manual reconfiguration for each new financial statement format.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4579567A1Machine learning based systems and methods for data extraction from financial statements
Publication Date: 2025.07.02 HIGHRADIUS CORP
  • EP4579567A1 patent drawingFigure 1
  • EP4579567A1 patent drawingFigure 2
  • EP4579567A1 patent drawingFigure 3

AI summary

A machine learning based computing method for extracting data from electronic documents, is disclosed. The machine learning based computing method includes steps of: receiving the electronic documents from data sources; converting formats of the electronic documents into first images associated with the electronic documents; classifying the first images associated with the electronic documents as second images including tabular data, and third images including non-tabular data, based on a CNN based machine learning model; determining tabular regions in tabular data associated with the second images by identifying bounding box coordinates indicating tabular regions in tabular data, based on a YOLO architecture based machine learning model; extracting first data from the tabular regions; extracting second data from non-tabular regions; and providing an output of the first data and the second data to second users on a user interface associated with electronic devices.