Variance-Stabilizing Transformation for Inspection Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
It is challenging to detect hidden objects in inspection images corrupted by noise, especially when the Signal-to-Noise Ratio (SNR) is small, as zooming in on these images exacerbates the noise issue, making it difficult to identify objects like weapons or dangerous materials.
Innovation Solution
The method involves applying a variance-stabilizing transformation to inspection images corrupted by Poisson-Gaussian noise, followed by denoising using a filter like Non-Local Means, and then optionally zooming in on the denoised image using a deconvolution-based technique to enhance object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If zooming is applied to inspection images corrupted by noise, then the image is enlarged to detect smaller objects, but the noise is also enlarged making object detection more difficult
Solution Approach 1:
The patent applies denoising processing to the inspection image before performing zooming operations. By removing noise in advance through algorithms such as Non-Local Means filtering, the subsequent zooming operation works on a cleaner image, preventing noise amplification while still enabling enhanced detection of smaller objects.
Solution Approach 2:
The patent transforms the noise characteristics by applying a variance-stabilizing transformation that converts Poisson-Gaussian noise into additive Gaussian noise with constant variance. This parameter change in noise characteristics enables more effective denoising and prevents noise amplification during zooming operations.
2Object-affected harmful factors
If denoising is applied to inspection images, then noise is reduced improving object visibility, but the processing time and computational complexity increase
Solution Approach 1:
The patent applies denoising processing selectively to regions of interest or to the entire image only when necessary, rather than applying it uniformly to all images. This partial application approach reduces computational time while still providing effective noise reduction where needed for object detection.
3Stability of the object's composition
If variance-stabilizing transformation is applied to transform Poisson-Gaussian noise, then noise variance becomes constant enabling better denoising, but the transformation complexity increases
Solution Approach 1:
The patent applies a variance-stabilizing transformation that specifically targets the noise characteristics by transforming Poisson-Gaussian noise into additive Gaussian noise with constant variance. This parameter change in noise statistics simplifies subsequent denoising operations and enables more effective processing.
Data Source
Figure 1
Figure 2
Figure 3A~4B
AI summary
In one aspect, there is provided a method of denoising one or more inspection images comprising a plurality of pixels, comprising: receiving an inspection image generated by an inspection system configured to inspect one or more containers, the inspection image being corrupted by a Poisson- Gaussian noise and a variance of the noise being non-constant in the plurality of pixels, and denoising the received inspection image by applying, to the inspection image, a variance-stabilizing transformation for transforming the variance of the noise into a constant variance in the plurality of pixels, wherein the variance-stabilizing transformation is based on a descriptor associated with the angular divergence of the inspection radiation and the Poisson-Gaussian noise.