Parallel Neural OCR for Check Data Field Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Optical Character Recognition (OCR) processes for financial documents, such as checks, suffer from high error rates, especially when processing millions of units, leading to inefficiencies and resource wastage due to errors like 3% resulting in over 30,000 errors, which are unsustainable for efficient document processing.
Innovation Solution
Implementing a set of virtual servers with customized neural networks, each dedicated to extracting specific data fields, utilizing parallel processing configurations and machine learning algorithms to enhance accuracy and efficiency in extracting data fields from financial documents, particularly checks, by training and tuning neural networks on historical or synthetic data representative of common check data types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single OCR process is used to extract all data fields from financial documents, then the system complexity is low, but the extraction accuracy deteriorates due to high error rates
Solution Approach 1:
The patent divides the data extraction task into separate specialized OCR processes, with each process dedicated to extracting specific data fields (e.g., one process for account numbers, another for dates). This segmentation allows each process to be optimized for its specific field type, thereby improving overall extraction accuracy while managing complexity through modular architecture.
Solution Approach 2:
The patent implements a universal framework that can handle multiple data field types through a single system architecture. The framework provides common functionality for image processing and field identification while allowing specialized extraction processes to be dynamically selected based on the specific field being processed, achieving both accuracy and system-wide efficiency.
2Measurement precision
If multiple images are captured to improve extraction accuracy, then the extraction accuracy improves, but the time consumption and resource usage increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing the captured image to identify and locate all data fields before extraction begins. The system determines the precise locations of account numbers, dates, and other fields in advance, allowing the specialized OCR processes to extract data directly from the single captured image without requiring multiple captures or sequential processing steps.
Solution Approach 2:
The patent creates a digital copy or representation of the financial document image that can be processed by multiple specialized OCR processes simultaneously. This copied representation allows parallel extraction of different data fields from the same source image, improving accuracy through multiple extraction attempts without requiring multiple physical captures or increasing time consumption.
3Productivity
If a single neural network model is used for all data fields, then the device complexity is low, but the extraction efficiency deteriorates due to processing all fields sequentially
Solution Approach 1:
The patent segments the data extraction function into separate specialized neural network processes, with each process trained specifically for its target data field. This segmentation enables parallel processing of different fields simultaneously, dramatically improving extraction efficiency and productivity while maintaining manageable complexity through modular process design.
Solution Approach 2:
The patent transitions from a single sequential processing dimension to a multi-dimensional parallel processing architecture. By organizing extraction processes across multiple dimensions (different specialized processes running simultaneously for different fields), the system achieves exponential improvement in productivity without linear increase in complexity.
4Measurement precision
If historical data is used to train neural networks, then the extraction accuracy improves, but the data processing volume and resource requirements increase
Solution Approach 1:
The patent applies local quality by training specialized neural networks on locally relevant historical data specific to each data field type. Rather than processing all historical data uniformly, the system selects and processes only the portion of historical data relevant to each specific field (e.g., historical account number formats, historical date patterns), thereby improving accuracy for each field while reducing overall processing volume and resource requirements.
Data Source
AI summary
A computer implemented method, system, and non-transitory computer-readable device that may be used in a remote deposit environment. A plurality of virtual servers, in a set of virtual servers, are assigned customized neural networks, such as convolutional neural networks (CNNs), based on an architecture and features of the data field, to extract, in parallel, data from specific data fields on a document. The selected customized neural networks are trained by historical or synthetic data corresponding to the data fields. Upon receiving, from a neural network optical character recognition (OCR) system, a selected first trained customized neural network model and at least a second selected trained customized neural network model, extract the data fields, based on parallel processing by the virtual servers, the customized neural networks, and the extracted data communicated to a remote deposit process.