Deep Learning OCR With Synthetic Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Optical Character Recognition (OCR) systems fail to achieve human-level accuracy required for automated business processes, particularly in handling documents with fancy fonts and patterned noise, and require expensive human labeling for training data generation.
Innovation Solution
A deep-learning based OCR system that employs a combination of synthetic and real-life text segments, using weak OCR systems to generate training data rapidly and reduce the need for human labeling, incorporating techniques like curriculum learning and statistical selection to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional OCR systems are used, then implementation is straightforward, but accuracy is insufficient for automated business processes
Solution Approach 1:
The patent segments the training data generation process into multiple components: weak OCR systems generate initial text, human reviewers correct errors, and synthetic documents are created from corrected text. This segmentation allows the system to achieve high accuracy without requiring all training data to be manually labeled, resolving the contradiction between reliability and complexity.
Solution Approach 2:
The patent applies preliminary action by using weak OCR systems to pre-process and generate initial text labels before human review. This preliminary step reduces the amount of manual labeling required while still achieving high accuracy, as human reviewers only need to correct errors rather than create labels from scratch.
2Reliability
If human labeling is used to generate training data, then accuracy is high, but time consumption and cost increase significantly
Solution Approach 1:
The patent uses synthetic document generation to create training data by copying and transforming corrected text into new document images. This copying approach allows the system to generate unlimited training examples from a limited set of corrected samples, dramatically reducing the time and cost of data preparation while maintaining high accuracy.
Solution Approach 2:
The system implements self-service by using the corrected text from human reviewers to automatically generate synthetic training documents. This self-service mechanism eliminates the need for continuous human labeling of all training data, as the system can generate additional training examples autonomously from the corrected samples.
3Reliability
If more training data is collected to improve accuracy, then model performance increases, but data collection and labeling costs increase
Solution Approach 1:
The patent applies copying by generating synthetic training documents from corrected text samples. This allows the system to create large volumes of training data without proportional increases in human labeling costs, as the synthetic generation process is automated and can produce unlimited examples from a small set of corrected samples.
Solution Approach 2:
The patent changes the parameter of training data source from exclusively real documents to a mixture of real and synthetic documents. This parameter change allows the system to achieve high model accuracy without collecting and labeling proportionally large amounts of real training data, as synthetic documents provide additional diverse examples at low cost.
Data Source
AI summary
An optical character recognition system employs a deep learning system that is trained to process a plurality of images within a particular domain to identify images representing text within each image and to convert the images representing text to textually encoded data. The deep learning system is trained with training data generated from a corpus of real-life text segments that are generated by a plurality of OCR modules. Each of the OCR modules produces a real-life image/text tuple, and at least some of the OCR modules produce a confidence value corresponding to each real-life image/text tuple. Each OCR module is characterized by a conversion accuracy substantially below a desired accuracy for an identified domain. Synthetically generated text segments are produced by programmatically converting text strings to a corresponding image where each text string and corresponding image form a synthetic image/text tuple.