Image Text Extraction with Rotation-Aware Attention for Noisy Documents
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in efficiently extracting text from unstructured scanned documents with noise and varying orientations, leading to increased handling time and inaccuracies in order processing due to skewed and noisy images.
Innovation Solution
A pipeline is developed using overlapping kernel-based embedding and Multi-head Context Attention in a customized neural network architecture to generate richer tokens, detect skewed text, and improve accuracy, combined with uniform scaling and rotation-sensitive regression to automate text extraction from images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional text extraction methods are used on unstructured scanned documents, then the extraction process is simple, but the accuracy is low due to noise and varying orientations
Solution Approach 1:
The patent segments the text extraction process into distinct modules: orientation detection module that identifies skewed text angles, extraction module that retrieves text based on detected orientations, and processing module that handles different orientation cases separately. This segmentation allows each module to specialize in specific tasks, improving overall accuracy while managing complexity through modular design.
Solution Approach 2:
The patent changes the parameter of text orientation from unnormalized (varying skew angles) to normalized (standardized orientations). By detecting the orientation angle of skewed text and transforming it to a standard orientation, the system can apply consistent extraction rules, significantly improving extraction accuracy for documents with varying orientations and noise.
2Productivity
If manual processing of skewed and noisy images is performed, then accuracy can be maintained, but the handling time increases significantly
Solution Approach 1:
The system implements self-service through automated orientation detection and text extraction. The orientation detection module automatically identifies skewed text angles without human intervention, and the extraction module autonomously retrieves text based on detected orientations. This automation eliminates manual processing steps while maintaining high accuracy, reducing average handling time from 10-12 minutes to significantly lower levels.
Solution Approach 2:
The patent applies preliminary action by detecting and correcting text orientation before the actual extraction process. The system first identifies the orientation angle of skewed text, performs normalization transformation, and then proceeds with text extraction. This preliminary orientation correction ensures that subsequent extraction operations can be performed efficiently and accurately without manual intervention.
3Productivity
If automated extraction systems are implemented, then handling time is reduced, but inaccuracies increase due to noise and skew in images
Solution Approach 1:
The patent applies local quality by treating different orientation regions of the document differently. The orientation detection module identifies specific skew angles in different regions, and the extraction module applies region-specific extraction rules. This allows the system to handle local variations in noise and skew patterns while maintaining high overall accuracy, rather than applying a uniform approach that would fail on varied documents.
Solution Approach 2:
The system implements feedback through iterative orientation detection and extraction validation. The orientation detection module continuously refines angle estimates based on detected text patterns, and the extraction module provides feedback on extraction quality. This feedback loop allows the automated system to adapt to noise and skew variations, maintaining high accuracy while achieving fast processing speeds.
Data Source
AI summary
A method for extracting text information from images includes obtaining an extraction request associated with live data comprising an image; generating, using a prediction model, rotational variant features and rotational invariant features associated with the live data; generating, using the prediction model, text embeddings associated with the rotational variant features using overlapping kernel-based embedding on the live data; generating, using the prediction model, attention values for each pixel in the live data using context attention; applying a trained language model to the text embeddings, attention values, and the live data to generate predictions; and performing extraction actions based on the predictions.