Document Title Judgment Using Selective Character Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for classifying documents scanned continuously face challenges such as converting separate documents into a single image and misclassifying continuous documents into different images per page, with labor-intensive solutions like inserting white sheets being necessary.
Innovation Solution
An information processing apparatus with a control device that extracts character areas, calculates feature quantities, generates a machine learning model for title judgment, calculates title reliability levels, performs character recognition, and judges titles based on pre-stored candidates, thereby automatically classifying documents without manual intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all character areas are converted to text data for title judgment, then title judgment accuracy is improved, but processing time and computational load increase significantly
Solution Approach 1:
The patent applies partial action by selectively converting only certain character areas to text data rather than all character areas. The system identifies and converts only those character areas with high title reliability scores (above a threshold) into text data for title judgment, while leaving other character areas unconverted. This selective approach maintains title judgment accuracy for relevant areas while significantly reducing overall processing time and computational load.
2Measurement precision
If machine learning model processes all character areas, then title detection accuracy is improved, but processing load increases
Solution Approach 1:
The system applies partial action by having the machine learning model process only character areas that meet specific criteria (high title reliability scores) rather than all character areas uniformly. This selective processing maintains title detection accuracy for relevant character areas while significantly reducing the overall processing load on the system.
Solution Approach 2:
The patent applies local quality by assigning different processing treatments to different character areas based on their individual title reliability scores. Character areas with high title reliability undergo full machine learning processing and text conversion, while character areas with low title reliability are excluded from intensive processing. This localized quality approach optimizes resource allocation while maintaining detection accuracy where it matters most.
Data Source
AI summary
In an image reading apparatus, a character area extractor extracts character areas from a document image in units of rows. A title reliability calculator calculates a title reliability level of each character area using a feature quantity data set and a machine learning model. A character recognizer converts character areas of which title reliability levels exceed a threshold into text data. A title judger collates text data with a title candidate. In a case in which one piece of text data coinciding with a title candidate is judged and detected, the title judger sets the text data as a title of the document image. In a case in which a plurality of pieces of coinciding text data are judged and detected, the title judger sets text data of which a title reliability level is the highest among the detected pieces of text data as a title of the document image.