Wide Spectrum Denoising for Microscopic Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for denoising microscopic images, particularly those from STORM technology, are inadequate for effectively addressing Poisson and mixed noise, which limits the temporal and spatial resolution and increases localization errors, especially when using high-resolution cameras.
Innovation Solution
A wide spectrum denoising method that involves extracting sub-block images, performing iterative optimization on a measurement matrix based on the point spread function, and using singular value decomposition to suppress noise, while maintaining image integrity by reshaping and splicing the denoised image, is developed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If high-resolution camera is used (effective pixel much smaller than PSF standard deviation), then spatial resolution is improved, but noise increases and localization accuracy decreases
Solution Approach 1:
The patent introduces compressed sensing as an intermediary method between high-resolution camera acquisition and final localization. The measurement matrix acts as a mediator that transforms the high-dimensional noisy data into a lower-dimensional space where noise is suppressed, then reconstructs the image to achieve both high spatial resolution and accurate localization.
Solution Approach 2:
The patent changes the parameter of camera pixel size relative to PSF standard deviation, using much smaller pixels than traditional requirements. Combined with compressed sensing reconstruction, this parameter change enables high spatial resolution while the denoising process maintains localization accuracy by separating signal from noise in the transformed domain.
2Object-affected harmful factors
If traditional denoising algorithms (BM3D, GAV) are applied to STORM raw images, then Gaussian or Poisson-Gaussian noise is reduced, but they are not suitable for the mixed noise characteristics of STORM images
Solution Approach 1:
The patent transforms the noise model parameters by applying Anscombe variance-stabilizing transformation to convert Poisson-distributed noise into approximately Gaussian noise with stable variance. This parameter transformation enables the use of Gaussian-based denoising methods while maintaining effectiveness for the original Poisson-Gaussian mixed noise characteristics of STORM images.
Solution Approach 2:
The patent introduces a specialized measurement matrix as an intermediary that is specifically designed for STORM image characteristics. This matrix serves as a bridge between the raw noisy image and the denoised output, incorporating knowledge of PSF and noise statistics to achieve optimal denoising performance tailored to STORM imaging.
3Object-affected harmful factors
If single molecule localization algorithms perform bandpass filter processing on raw images, then some noise is reduced, but a lot of information is lost which is not suitable for CS reconstruction
Solution Approach 1:
Instead of filtering the raw image before reconstruction (the conventional approach), the patent inverts the order by first performing compressed sensing reconstruction on the raw image, then applying denoising to the reconstructed image. This reversal preserves all original information during reconstruction while removing noise afterward, avoiding information loss.
Data Source
AI summary
The present invention discloses a wide spectrum denoising method for microscopic images, comprising: connecting a sub-block image matrix end to end to convert same into a one-dimensional vector yraw; performing iterative optimization processing on a measurement matrix A to obtain an optimization matrix Ao; calculating a transition matrix T based on the measurement matrix A and the optimization matrix Ao, and performing singular value decomposition on the transition matrix T to obtain USVT; compressing the value greater than the threshold in SVTyraw to the threshold, and unchanging the value less than the threshold, thus achieving the purpose of denoising; finally, left multiplying the noise-suppressed y′sv by T−1U to obtain denoised YWSD, then cutting off overlapping parts of edges, and splicing a complete denoised image row by row or column by column.