Document Image Text Recognition Using Connected Component Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR systems are unsuitable for document images with poor quality and hand-filled forms due to poor text recognition, requiring significant user input and computing resources, and struggle with identifying fields and regions of interest across varying document templates and image quality.
Innovation Solution
A computer-implemented method that identifies connected components in images, generates vector representations, and uses object detection models for supervised training to extract matching text and identify fields and regions of interest, enabling accurate text extraction and image data retrieval from document images with poor quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing OCR systems are used for document images with poor quality and hand-filled forms, then text recognition accuracy deteriorates, but computing resources and user input requirements increase
Solution Approach 1:
The patent segments the document image into multiple connected components (text regions, fields, and regions of interest) using image processing techniques. This segmentation allows the system to process and analyze different parts of the document independently, improving text recognition accuracy in poor quality images by focusing on specific connected components rather than processing the entire image at once.
Solution Approach 2:
The patent introduces vector representations as an intermediary between the image data and the object detection model. These vector representations capture the essential features of connected components and enable the model to learn patterns and make accurate predictions about text locations and field boundaries, thereby improving recognition accuracy without requiring excessive computing resources or user input.
2Measurement precision
If existing OCR systems process document images, then text extraction is performed, but fields and regions of interest cannot be reliably identified across varying document templates
Solution Approach 1:
The patent creates a universal object detection model that can identify multiple types of elements (text, fields, regions of interest) across different document templates using a single trained system. The model is trained to recognize common patterns and relationships between connected components, enabling it to adapt to varying document structures without requiring template-specific processing or user input for each document type.
Solution Approach 2:
The patent employs feedback mechanisms where the system analyzes the relationships between connected components and uses this information to refine its identification of fields and regions of interest. The object detection model continuously learns from the spatial relationships and contextual information present in the document images, improving its ability to accurately identify fields across different templates.
3Measurement precision
If supervised training is applied to improve text recognition, then accuracy improves, but training data requirements and system complexity increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing the document images to identify connected components and generate vector representations before feeding the data to the object detection model. This preliminary processing reduces the complexity of the training process by preparing the data in a structured format that the model can more easily learn from, thereby improving recognition accuracy without requiring excessively large or complex training datasets.
Data Source
AI summary
Systems and methods are configured to extract text from images (e.g., document images) utilizing a combination of optical character recognition processes and neural network-based analysis of various images to train a machine-learning object recognition model that is configured to identify text within images based on object-comparisons between known and unknown text within images. In certain embodiments, identified text within images may be utilized to identify corresponding regions-of-interest for extraction of image data encompassing a portion of an image that may be indexed based at least in part on text identified as corresponding to the particular region-of-interest.