AI Image Normalization for Cross-Vendor Processing Reliability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based image processing systems, such as convolutional neural networks (CNNs), perform reliably only if the input images are within the distribution range of images used for training, and are not effective when applied to images from different vendors or unknown sources due to varying image characteristics.
Innovation Solution
A method that determines an image metric for input images, generates a filter to transform the image to match target values, and applies standardized image processing methodologies, such as AI-based algorithms, while remaining agnostic to the image source, using techniques like modulation transfer function (MTF) and noise power spectrum (NPS) adjustments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If AI-based image processing is applied to images from different vendors, then image processing versatility is improved, but processing reliability deteriorates due to varying image characteristics outside the training distribution
Solution Approach 1:
The patent introduces an intermediary transformation process that converts images from different vendors into a standardized representation that matches the training distribution of the AI model. This intermediary step (filter generation and application) acts as a bridge between diverse input images and the fixed requirements of the neural network, enabling versatile processing while maintaining reliability.
Solution Approach 2:
The system dynamically adjusts image parameters by generating custom filters based on the statistical characteristics of input images from different vendors. These parameter transformations (normalization, filtering) adapt the input images to match the expected training distribution, resolving the contradiction between handling diverse inputs and maintaining consistent processing quality.
2Productivity
If standardized reconstruction filters are used for AI denoising, then processing efficiency is improved, but adaptability to different image sources deteriorates
Solution Approach 1:
The system transitions from static standardized filters to dynamic, adaptive filters that are generated on-the-fly based on the statistical properties of each input image. This dynamic filter generation maintains processing efficiency while significantly improving adaptability to different image sources and characteristics.
Solution Approach 2:
The system performs preliminary analysis of input image characteristics (noise power spectrum, modulation transfer function) before applying the main AI processing. This preliminary action enables the generation of optimized filters tailored to each specific input, improving both efficiency and adaptability by avoiding trial-and-error approaches.
3Measurement precision
If neural networks are trained on narrow distribution of images, then processing precision is improved, but versatility across different image types deteriorates
Solution Approach 1:
Instead of training the neural network on diverse images to improve versatility (which would compromise precision), the patent inverts the approach by keeping the network trained on a narrow, high-quality distribution for precision, and handling versatility through pre-processing transformations that adapt diverse inputs to match the training distribution.
Data Source
AI summary
Images are processed while remaining agnostic as to an image source. An input image to be processed is retrieved. A first value or set of values for an image metric associated with the input image is determined, and a first filter is generated based on a relationship between the first value or set of values and a target value or set of values. The first filter is then applied to the input image to generate a working image having a second value or set of values for the image metric substantially similar to the target value or set of values for the image metric. The working image is processed using a standardized image processing methodology. An output is generated, which may be an image, based on the processed working image and outputs the output.