Machine Learning Filter Selection for Document Image Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The capture of physical documents into digital files can be imperfect due to issues like improper alignment and insufficient lighting, leading to blurry images that may not allow for accurate determination of transaction data, resulting in inefficient processes and user frustration.
Innovation Solution
A machine-learned model is used to automatically apply filters to images by generating hash patterns and mapping them to filters, selecting the most accurate filter based on Hamming distance, and applying it to improve image quality for data extraction using OCR.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image capture is performed under suboptimal lighting conditions or with improper alignment, then the capture process is simple and quick, but the image quality becomes poor leading to inaccurate data extraction
Solution Approach 1:
The system performs preliminary actions by applying multiple filters (sharpness filter, brightness filter, contrast filter) to the captured image before data extraction. These filters are applied in advance to compensate for suboptimal capture conditions, transforming a potentially poor quality image into an improved version that enables accurate data extraction without requiring the user to recapture the image.
Solution Approach 2:
The system changes image parameters by adjusting sharpness, brightness, and contrast through filter application. The machine learning model determines which filters to apply and at what intensities based on the captured image's characteristics, thereby improving image quality parameters to enable accurate OCR while maintaining the simplicity of the capture process itself.
2Measurement precision
If multiple filters are applied to improve image quality, then data extraction accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The system applies local quality by selectively applying filters based on the specific defects detected in the captured image. Instead of always applying all possible filters, the machine learning model analyzes the image and applies only the necessary filters (sharpness, brightness, or contrast) to address the specific issues present, thereby improving accuracy while minimizing unnecessary processing time.
Solution Approach 2:
The system uses feedback from the machine learning model that analyzes the captured image's quality characteristics to determine which filters should be applied. This feedback mechanism allows the system to intelligently select and apply only the necessary filters, avoiding unnecessary computational resources while ensuring accurate data extraction when needed.
3Measurement precision
If the system requires perfect image capture to ensure accurate data extraction, then data extraction accuracy is maintained, but user experience deteriorates due to repeated capture attempts
Solution Approach 1:
The system converts the harm of suboptimal capture conditions into benefit by using machine learning to detect issues like poor lighting, blurriness, or low contrast, and automatically applying corrective filters. This transforms what would normally be a rejection requiring recapture into an opportunity to improve the image and successfully extract data, thereby maintaining accuracy while improving user experience.
Solution Approach 2:
The system performs self-service by automatically detecting image quality issues and applying appropriate filters without requiring user intervention. The machine learning model autonomously analyzes the captured image, determines necessary corrections, and applies filters to improve quality, allowing users to simply capture the image without needing to manually adjust lighting or alignment conditions.
Data Source
AI summary
Implementations of the present disclosure include receiving a training image, providing a hash pattern that is representative of the training image, applying a plurality of filters to the training image to provide a respective plurality of filtered training images, identifying a filter to be associated with the hash pattern based on the plurality of filtered training images, and storing a mapping of the filter to the hash pattern within a set of mapping in a data store.