Document Field Recognition via Location Transformations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in efficiently grouping and classifying scanned documents with optical character recognition (OCR) errors, text differences, graphics, noise, rotations, skewing, and varying formats, which complicates the transformation of paper documents into intelligent content for enterprise systems.
Innovation Solution
The method employs location transformations to identify objects by generating word pairs, computing location information, and applying transformations such as translation, rotation, and scaling to align anchor positions across documents, thereby improving field recognition rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR methods are used to process scanned documents, then text extraction can be performed, but accuracy deteriorates due to OCR errors, text differences, graphics, noise, rotations, and skewing
Solution Approach 1:
The patent segments the document processing task into multiple stages: generating word pairs from OCR data, computing location information for each word, identifying transformations between documents, and applying transformations to align anchor positions. This segmentation allows each stage to address specific issues (OCR errors, positioning variations, scaling) independently, improving overall field recognition accuracy despite individual stage limitations
Solution Approach 2:
The patent changes the parameter space by introducing transformation parameters (translation, rotation, scaling) that describe the relationship between documents. By modeling document variations as parameter changes rather than absolute position mismatches, the system can accommodate rotations, skewing, and scaling while maintaining accurate field recognition through relative position comparisons
2Measurement precision
If location transformations are applied to handle document variations, then field recognition accuracy improves, but computational complexity increases due to generating word pairs, computing locations, and applying transformations
Solution Approach 1:
The patent performs preliminary actions by generating word pairs and computing location information before actual document comparison. By pre-processing OCR data to extract word locations and relationships, the system prepares structured information that simplifies subsequent transformation applications and anchor identification, reducing the complexity of the main comparison task
Solution Approach 2:
The patent introduces transformation models as intermediaries between raw OCR data and final field recognition results. The transformation parameters act as a mediator that bridges the gap between documents with different positioning, scaling, and orientation, allowing the system to handle variations without directly comparing every element across all documents
3Adaptability or versatility
If documents are scanned in various positions and orientations, then document collection versatility improves, but processing difficulty increases due to rotations, skewing, and positioning variations
Solution Approach 1:
The patent models document positioning variations as parameter changes (translation, rotation, scaling) rather than fundamental differences. By representing each document's position in terms of transformation parameters relative to a reference document, the system can accept documents in any orientation while maintaining consistent anchor identification through parameter adjustment
Solution Approach 2:
The patent adds a transformation parameter dimension to the document representation. Instead of working only with absolute pixel coordinates, the system introduces relative position parameters that describe how one document transforms into another, effectively adding a layer of abstraction that handles rotations and skewing by treating them as parameter variations rather than positioning errors
Data Source
AI summary
Using location transformations to identify objects is described. Word pairs are generated. Each word pair includes a first word from a first document and a corresponding second word from a second document. For each word pair, location information is computed for the words that indicates locations of the words in the documents relative to other words in the documents. A transformation is identified based on a comparison between the first and second location information. The transformation includes a translation, a rotation, and/or a scale. The transformation is applied to the second location information. A first anchor is identified in the first document if a difference between the first location information and the transformed second location information is less than a threshold value. A second anchor is identified in the second document based on the first anchor and the transformation. A field may be identified in the second document based on the second anchor and at least partially based on the transformation.