ML Image Generation Reliability via Statistical Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for solving inverse problems, such as image generation in seismic imaging, often result in images with artifacts, and increasing data quantity to reduce artifacts is costly and resource-intensive.
Innovation Solution
The use of machine-learning models to generate multiple sets of image data based on different waveform return measurements and model parameters, allowing for the determination of a representative image and the identification of reliable areas within the image through statistical evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If optimization techniques are used to generate solutions to inverse problems, then image data can be generated from waveform return measurements, but artifacts are introduced in the generated images
Solution Approach 1:
The patent segments the image generation process into multiple independent trials, each producing a candidate solution. By dividing the problem into multiple separate optimization runs with different initial conditions and model parameters, the system can later evaluate and select the most reliable solution while filtering out artifacts through statistical analysis of results across trials.
Solution Approach 2:
The patent implements feedback through statistical evaluation of multiple candidate solutions. By comparing results across multiple trials and using statistical metrics to assess consistency and reliability, the system provides feedback that identifies and eliminates artifact-contaminated solutions, thereby improving overall image accuracy.
2Reliability
If the quantity of data is increased to reduce artifacts, then image quality improves, but data collection cost and time increase
Solution Approach 1:
The patent creates multiple copies of the image generation process through repeated trials rather than increasing the quantity of input data. Each trial generates a candidate solution that serves as a copy, and statistical evaluation of these copies identifies the most reliable result without requiring additional waveform return measurements.
Solution Approach 2:
The patent changes model parameters and initial conditions across multiple trials instead of increasing data quantity. By varying parameters such as optimization starting points and model configurations, the system generates diverse candidate solutions that can be statistically evaluated to produce high-quality images without collecting more waveform return data.
3Reliability
If the quantity of data is increased to reduce artifacts, then image quality improves, but computing resources required increase dramatically
Solution Approach 1:
The patent applies partial action by performing multiple optimization trials with moderate data sets rather than one exhaustive optimization with complete data. Each trial uses a portion of the available computational resources, and statistical evaluation selects the best result, achieving high image quality without requiring all computational resources to be consumed by a single large-scale optimization.
Data Source
AI summary
A method includes using a machine-learning model to determine multiple sets of image data, each representing an estimated solution to an inverse problem associated with multiple waveform return measurements. First image data are based on a first set of waveform return measurements and first model parameters of the machine-learning model, and second image data are based on a second set of waveform return measurements and a second model parameters of the machine-learning model. The method also includes determining, based on the multiple sets of image data, a representative image. The method further includes generating output data that identifies a first area of the representative image as less reliable than a second area of the representative image based on a statistical evaluation of two or more sets of image data of the multiple sets of image data.