Collaborative Image Filtering via Frequency Domain Sharpening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional self-similarity based techniques for noise reduction and image enhancement fail when there are low spatial correlations among image patches, leading to loss of important details.
Innovation Solution
A two-stage collaborative filtering and sharpening process that involves constructing 3D stacks of image blocks, applying 1D transforms, modifying transform-domain coefficients through shrinking and alpha-rooting, and using Wiener filtering to enhance image sharpness and reduce noise, even in images with low spatial correlations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If conventional self-similarity based techniques are applied to noise reduction, then noise is reduced effectively when high spatial correlations exist among image patches, but important details are lost when spatial correlations are low
Solution Approach 1:
The patent transforms the filtering problem from spatial domain to frequency domain by applying 1D Fourier transform along the stacking dimension of 3D stacks. This dimensional transformation allows the filter to operate in a different domain where noise and signal can be separated more effectively, avoiding the loss of details that occurs in spatial domain filtering when spatial correlations are low.
Solution Approach 2:
The patent modifies filter parameters dynamically by computing frequency response characteristics and adjusting the frequency response of the 1D Fourier transform filter based on the estimated power spectral density of the image. This adaptive parameter adjustment allows the filter to optimize noise reduction while preserving details according to the specific characteristics of each image region.
2Object-affected harmful factors
If aggressive filtering is applied to reduce noise, then noise reduction performance improves, but image sharpness and detail preservation deteriorate
Solution Approach 1:
The patent applies different filtering strategies to different frequency components locally. By analyzing the power spectral density and computing frequency response characteristics for each 3D stack, the filter adapts its strength locally - applying stronger filtering to noise-dominated frequency components while preserving signal-rich components, thus maintaining image sharpness while reducing noise.
Solution Approach 2:
The patent employs an iterative refinement process where the filtered output is used to update the estimate of the clean image's power spectral density, which in turn updates the frequency response of the filter. This feedback mechanism allows the filter to progressively improve noise reduction while preserving details, balancing noise reduction and sharpness through iterative optimization.
3Object-affected harmful factors
If collaborative filtering is applied to enhance image quality, then noise reduction improves, but computational complexity increases due to 3D stack construction and transform operations
Solution Approach 1:
The patent divides the image into multiple overlapping 2D patches, which are then stacked to form 3D stacks. This segmentation allows the filtering operation to be performed independently on each 3D stack in the frequency domain, parallelizing the computation and reducing overall complexity compared to processing the entire image at once.
Solution Approach 2:
The patent replaces complex spatial domain filtering operations with simpler frequency domain operations. By applying 1D Fourier transform along the stacking dimension and performing multiplication in the frequency domain, the computational complexity is reduced compared to traditional spatial domain collaborative filtering that would require complex convolution operations.
Data Source
AI summary
Various techniques are disclosed for reducing noise and enhancing sharpness of an input image. For example, a method includes performing an initial collaborative filtering and sharpening on the input image to generate a pilot image, using the pilot image to derive coefficients that are used to perform a second collaborative filtering on the input image to generate a filtered image. In some embodiments, the collaborative filtering and sharpening is performed using parameters that boost or enhance the differences in pixel values for the same spatial locations of the matched image blocks extracted during the collaborative filtering and sharpening process. Accordingly, the method according to various embodiments performs especially well for images that have weak spatial correlations among mutually similar blocks.