Form Field Identification via Image Segmentation and ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Optical character recognition (OCR) techniques fail to accurately identify fields and labels in digital images of forms, especially when image quality is poor or when labels are in non-standard fonts or unusual formats, leading to incorrect interpretations and loss of context.
Innovation Solution
The use of computer-vision image-segmentation techniques to divide images into segments, combined with machine-learning models that classify these segments without relying on OCR, allowing for the detection and quantification of features and the classification of image segments as fields or labels, even in the absence of a template.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If OCR is applied to digitize text on paper forms, then text can be converted into computer-encoded text, but accuracy deteriorates when image quality is poor or labels are in non-standard fonts and unusual formats
Solution Approach 1:
The patent divides the form image into multiple segments based on spatial relationships and visual features. Each segment is then processed independently to identify labels and fields, allowing the system to handle complex layouts and unusual formatting more effectively than global OCR approaches.
Solution Approach 2:
The system changes the approach from character recognition to feature-based classification. By extracting visual features from image segments and using machine learning models to classify them as labels or fields, the system bypasses the limitations of traditional OCR regarding font variations and image quality.
2Loss of information
If OCR is used to identify labels and fields, then text extraction can be performed, but context is lost and incorrect interpretations occur
Solution Approach 1:
The system performs preliminary segmentation and classification of image regions before attempting any text extraction. By first identifying which regions are labels and which are fields based on spatial relationships and visual features, the system preserves contextual information that would otherwise be lost in pure OCR approaches.
Solution Approach 2:
The patent introduces an intermediate classification step between image input and text extraction. Machine learning models classify image segments as labels or fields, serving as an intermediary that preserves contextual understanding before any character recognition occurs, thereby maintaining reliability.
3Measurement precision
If multiple segmentation methods are used to improve accuracy, then field and label identification improves, but processing time increases
Solution Approach 1:
The patent combines multiple segmentation approaches and integrates them into a unified processing pipeline. By merging the strengths of different segmentation methods and processing segments in parallel, the system achieves high accuracy while managing processing time efficiently through coordinated multi-method analysis.
Data Source
AI summary
Systems of the present disclosure allow fields and labels to be identified in a digital image of a form without performing OCR. A digital image of a form can be partitioned into image segments using computer-vision image-segmentation techniques. Features for each image segment can be extracted using computer-vision feature-detection methods. The features extracted from an image segment can be included in an input instance for a machine-learning model. The machine-learning model can assign a classification to the input instance. The classification can associate the input instance with a field type or a label type.