Synthetic Document Image Augmentation for Deep Learning Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated recognition of fields and associated values in business documents with varying formats and image quality is challenging due to scanning artifacts and limited availability of diverse document datasets for training machine learning systems.
Innovation Solution
A computerized system that synthetically generates distorted document images using augmentation modules to simulate scanning effects, allowing for precise training of deep learning networks and testing of preprocessing components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning systems are trained on real scanned document images, then the system learns to handle real-world scanning artifacts and distortions, but obtaining large diverse datasets is challenging due to confidentiality and cost
Solution Approach 1:
The patent creates synthetic copies of document images by applying various distortion transformations (skew, perspective, noise, compression artifacts) to clean template documents. This allows generating large volumes of training data without needing actual scanned documents, resolving the contradiction between data quality and data quantity
Solution Approach 2:
The system varies multiple parameters in the synthetic image generation process including distortion strength, noise levels, compression ratios, and transformation types. By systematically changing these parameters, the system generates diverse training samples that cover the range of real-world variations without requiring additional physical documents
2Reliability
If traditional computer vision approaches are used for preprocessing, then the system can handle document distortions, but the approach is less effective compared to deep learning methods
Solution Approach 1:
The patent applies preprocessing distortions to the training data before the deep learning model is trained. This preliminary action allows the model to learn distortion correction capabilities directly during training, eliminating the need for separate traditional computer vision preprocessing steps and achieving better accuracy
3Measurement precision
If deep learning systems are trained on diverse document formats and qualities, then recognition accuracy improves, but the variability in formatting and image quality makes automated processing challenging
Solution Approach 1:
The system generates synthetic training data with controlled variations in formatting parameters, image quality parameters, and distortion parameters. This systematic parameter variation allows the model to learn robust recognition across diverse formats while maintaining manageable system complexity through automated synthetic data generation
Data Source
AI summary
A computerized method and system for adding distortions to a computer-generated image of a document stored in an image file. An original computer-generated image file is selected and is processed to generate one or more distorted image files for each original computer-generated image file by selecting one or more augmentation modules from a set of augmentation modules to form an augmentation sub-system. The original computer-generated image file is processed with the augmentation sub-system to generate an augmented image file by altering the original computer-generated image file to add distortions that simulate distortions introduced during scanning of a paper-based representation of a document represented in the original computer-generated image file.