Multiscale Image Filtering via Hierarchical Kernel Approximation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing methods for applying large linear translation-invariant filters are computationally expensive, particularly due to the high cost of convolution operations, and lack translation invariance, which affects accuracy in tasks like gradient field integration and scattered data interpolation.

Innovation Solution

A multiscale scheme that uses numerically optimized small kernels to approximate the effect of large filters, allowing for efficient computation in linear time by repeatedly performing convolutions with fixed-width kernels while downsampling and upsampling images, tailored to directly reproduce the effect of specific LTI operators.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If convolution with large kernels is used for image processing, then filtering accuracy is improved, but computational cost increases to O(n2)

Engineering Contradiction:
Improvefiltering accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the large kernel convolution operation into multiple smaller kernel convolutions performed at different scales. By decomposing the filtering task into hierarchical levels with progressively smaller kernels, the method achieves O(n) computational complexity while maintaining accurate approximation of the original large kernel filtering effect.

Inventive Principle:
Principle #1Segmentation

2Productivity

If FFT-based methods are used for periodic domains, then computational cost reduces to O(nlogn), but translation invariance is lost due to subsampling operations

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidtranslation invariance
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies local quality by using fixed-width kernels at each scale that maintain consistent filtering properties regardless of position in the image. This local consistency ensures translation invariance is preserved, as the same kernel operations are applied uniformly across all regions and scales, unlike subsampling methods that introduce position-dependent artifacts.

Inventive Principle:
Principle #3Local quality

3Productivity

If separable kernel approximation is used, then computational cost reduces to O(kn), but accuracy deteriorates for non-separable kernels

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidapproximation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent transitions from 2D kernel approximation to a multi-scale hierarchical approach, adding the scale dimension to the problem. By operating at multiple resolutions and combining results, the method achieves accurate approximation of non-separable kernels without being constrained to separable forms, effectively solving the accuracy-efficiency trade-off through dimensional expansion.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP2724255B1Applying rapid numerical approximation of convolutions with filters for image processing purposes
Publication Date: 2020.12.09 YISSUM RESEARCH DEVELOPMENT COMPANY OF THE HEBREW UNIVERSITY OF JERUSALEM LTD
  • EP2724255B1 patent drawingFigure 1~2
  • EP2724255B1 patent drawingFigure 3A~3G
  • EP2724255B1 patent drawingFigure 4A~4F

AI summary

A computer implemented method for A computer implemented method for applying a numerical approximation of a convolution of image I as represented by hierarchical signals a l of level l with filter f, said method including the steps of: generating a forward transformation by applying a convolution between a l and kernel hI or each level of hierarchy l and by down-sampling the result of convolved a l and hI; generating a backward transformation by applying, for each level of hierarchy l, a convolution between kernel h 2 and an up-sampled forward transformation and combining the result with a convolution of al with kernel g; and combining the forward transformation with the backward transformation, to yield â 0 being an approximation of a convolution between a and f, wherein kernels hI, hI, and g are optimized kernels of filter f.