Polarization Model Estimation Using Noise Variance Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing polarization models calculated from unsaturated images are not robust against noise, particularly when the variance of observation values changes with luminance, leading to decreased accuracy.
Innovation Solution
An information generation device and method that calculates the amount of noise for each polarization direction using a noise variance model, allowing for the estimation of a polarization model that is robust against noise by correcting observation values and excluding saturated pixels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a high-sensitivity polarization model is calculated from an unsaturated image in a high-sensitivity polarization image, then polarization characteristics can be obtained with high sensitivity, but the accuracy of the calculated polarization model decreases when shot noise variance changes depending on luminance
Solution Approach 1:
The patent applies parameter changes by transitioning from uniform weighting in standard least squares fitting to non-uniform weighting based on noise variance. The noise variance model parameters (sigma squared) are calculated from image data and used to weight the polarization model fitting, adjusting the influence of each pixel based on its noise characteristics. This resolves the contradiction by making the fitting process adaptive to noise conditions rather than assuming uniform noise across the image.
Solution Approach 2:
The patent implements feedback by using the noise variance model to guide the polarization model calculation. The noise variance is estimated from the image data itself (using flat areas or still areas as references), and this estimated variance is then fed back into the fitting process to weight the observation values. This creates a self-adjusting system that improves polarization model accuracy while accounting for shot noise variations.
2Measurement precision
If standard least squares fitting is used without noise weighting, then the fitting process is simple and fast, but the accuracy deteriorates when noise variance is not uniform
Solution Approach 1:
The patent modifies the least squares fitting by introducing noise variance parameters as weighting factors. Instead of simple sum-of-squares minimization, the system performs weighted least squares where each observation value is weighted by the inverse of its noise variance. This parameter change allows the system to handle non-uniform noise while maintaining the mathematical simplicity and computational efficiency of least squares methods.
Solution Approach 2:
The patent applies preliminary action by pre-calculating the noise variance model from the image data before performing the polarization model fitting. The noise variance is estimated from flat areas or still areas in the image, and this pre-computed variance information is then used in the subsequent fitting process. This separation of noise estimation and model fitting simplifies the overall computation while improving accuracy.
3Measurement precision
If all observation values are used in the polarization model calculation, then the calculation uses maximum data, but saturated pixels reduce the accuracy of the polarization model
Solution Approach 1:
The patent applies the extraction principle by identifying and removing saturated pixels from the set of observation values used in polarization model calculation. The system detects saturated pixels (where the image sensor output is at maximum capacity) and excludes them from the fitting process. This ensures that only reliable, unsaturated measurement data contributes to the polarization model, preventing accuracy degradation from saturated regions.
Solution Approach 2:
The patent implements local quality by applying different treatment to different regions of the image based on their saturation state. Unsaturated pixels are used for polarization model calculation, while saturated pixels are excluded. This local differentiation ensures that each pixel's contribution to the model is appropriate for its quality, improving overall model accuracy by preventing corrupted data from degrading the results.
Data Source
AI summary
An information generation unit 30 acquires, from a polarization imaging unit 20, observation values in which polarization directions are at least three or more directions (m≥3). A noise amount calculation unit 35-1 calculates an amount of noise on the basis of an observation value in a first polarization direction. Similarly, noise amount calculation units 35-2 to 35-m calculate amounts of noise on the basis of observation values in second to m-th polarization directions. A polarization model estimation unit 36 estimates a polarization model by using the observation values for the respective polarization directions and the amounts of noise calculated by the noise amount calculation units 35-1 to 35-m. Thus, it is possible to calculate a polarization model that is robust against noise.