Document Image Text Classification for Expense Data Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The proliferation of images from cameras and electronic devices has led to challenges in efficiently extracting and processing data from images of documents, such as receipts, for applications like expense management.
Innovation Solution
A system that uses a machine learning model to detect and classify text in images of documents, determining bounding boxes for each text type, and generating sets of defined objects based on predicted text types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional data extraction methods are used on document images, then the process requires manual intervention and multiple processing steps, but the productivity and data extraction efficiency are low
Solution Approach 1:
The patent replaces manual mechanical data extraction processes with an automated machine learning-based system. The machine learning model automatically detects text regions, classifies text types, and extracts structured data from document images without human intervention, substituting the mechanical manual processing system with an intelligent automated system that significantly improves productivity and reduces processing time
Solution Approach 2:
The system enables self-service data extraction where the machine learning model autonomously performs text detection, classification, and extraction tasks. The model independently identifies different text types (e.g., dates, amounts, descriptions), determines their spatial relationships, and generates structured output without requiring manual guidance or intervention, allowing the system to serve itself in completing the entire data extraction workflow
2Measurement precision
If all detected text is processed equally, then the system maintains simplicity, but the accuracy of extracting specific defined objects is reduced
Solution Approach 1:
The patent segments the text processing task by classifying detected text into different types (e.g., dates, amounts, descriptions, vendor names) based on their content and spatial characteristics. Each text region is analyzed independently to determine its type, and then grouped with other text of the same type to form structured data objects. This segmentation approach improves classification accuracy by treating different text types with appropriate processing logic while managing complexity through modular classification rules
Solution Approach 2:
The system applies local quality by using different processing strategies for different text types based on their specific characteristics. For example, date text is validated against date formats, amount text is checked for numerical patterns, and description text is processed differently from vendor names. This localized processing approach tailors the extraction logic to the specific requirements of each text type, improving overall accuracy while maintaining manageable complexity through specialized handling rules
3Measurement precision
If a machine learning model is implemented for text classification, then the accuracy of object extraction is improved, but the device complexity and computational requirements increase
Solution Approach 1:
The patent implements preliminary action by first performing optical character recognition (OCR) to detect and transcribe all text in the document image before applying machine learning-based classification. The text detection phase identifies text regions and their spatial positions, creating a structured input for the subsequent classification stage. This preliminary text detection prepares the data in advance, allowing the machine learning model to focus specifically on classifying already-detected text rather than simultaneously detecting and classifying, thereby improving accuracy while managing computational complexity through staged processing
Data Source
AI summary
Some embodiments provide a non-transitory machine-readable medium that stores a program. The program receives an image of a document. The program further detects a plurality of text based on the image of the document. The program also uses a machine learning model to predict whether each text in the plurality of text is one of a plurality of defined types of text. Based on the predicted types of text for the plurality of text, the program further determines a set of defined objects.