Keypoint Descriptor Generation via Complex Wavelet Analysis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current image processing methods, such as SIFT and wavelet-based techniques, face challenges in accurately identifying objects in video sequences due to variations in bit rate, encoding methods, resolution, and color balance, which affect keypoint motion estimation across frames, and are not robust to transformations like rotation and translation.

Innovation Solution

A method that generates a keypoint descriptor invariant to transformations by using modified wavelet functions based on basis functions derived from object data, allowing for accurate object identification in images and video sequences, with reduced computational processing requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If SIFT method uses histograms for orientation assignment and descriptor generation, then rotation invariance is achieved, but processing complexity increases significantly

Engineering Contradiction:
Improverotation invarianceVSAvoidprocessing complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent replaces the histogram-based mechanical computation system with a complex wavelet transform system. Instead of computing histograms of gradient orientations, the method uses complex wavelets whose phases directly encode orientation information, and whose magnitudes provide descriptor values. This substitution of the computational mechanism reduces processing complexity while maintaining rotation invariance through the inherent phase properties of complex wavelets.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the fundamental parameters used for descriptor generation. Rather than using histogram bin counts and weighted averages, the method transforms image patches into the complex wavelet domain, where the real and imaginary components (or magnitude and phase) directly provide rotation-invariant descriptor values. This parameter change from histogram statistics to wavelet coefficients simplifies the computation while preserving rotational robustness.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If PCA-SIFT uses principal component analysis instead of histograms, then processing efficiency improves, but rotation invariance may be compromised

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidrotation invariance
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent replaces PCA's linear algebra-based dimensionality reduction mechanism with complex wavelet transform. Instead of computing covariance matrices and eigenvectors, the method directly transforms image patches into complex wavelet coefficients that inherently encode both orientation and magnitude information. This substitution maintains processing efficiency while ensuring rotation invariance through the phase properties of complex wavelets.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent uses complex wavelets as a composite mathematical tool that combines multiple functions into a single transform. The complex wavelet simultaneously provides orientation information through its phase, magnitude information through its amplitude, and rotation invariance through its complex structure. This composite approach replaces the multi-step PCA process with a unified transformation that achieves both efficiency and reliability.

Inventive Principle:
Principle #40Composite materials

3Measurement precision

If GLOH uses more spatial regions in histograms, then descriptor distinctiveness improves, but dimensionality and processing load increase

Engineering Contradiction:
Improvedescriptor distinctivenessVSAvoiddimensionality
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces GLOH's multi-region histogram computation with a single complex wavelet transform applied to the entire image patch. Instead of dividing the patch into multiple spatial regions and computing separate histograms, the complex wavelet transform processes the whole patch uniformly, with the spatial information naturally encoded in the wavelet coefficient magnitudes and the orientation information encoded in the phases. This reduces dimensionality while maintaining distinctiveness.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent transitions from the real-valued histogram space to the complex wavelet domain, adding the dimension of phase information. This complex dimension allows the descriptor to encode both magnitude and orientation simultaneously, reducing the need for multiple spatial regions while maintaining or improving descriptor distinctiveness. The complex plane provides a more efficient representation space than traditional histogram bins.

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

4Ease of operation

If video sequences are processed frame by frame independently, then processing simplicity is maintained, but object tracking accuracy decreases due to transformations

Engineering Contradiction:
Improveprocessing simplicityVSAvoidobject tracking accuracy
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent applies complex wavelet transform to extract rotation-invariant and translation-invariant descriptors from each frame before performing matching or tracking operations. By preprocessing the frames with this invariant descriptor extraction, the method prepares the data in advance to be robust against transformations, allowing simple frame-by-frame processing to achieve accurate object tracking despite rotations and translations between frames.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2353126B1Keypoint descriptor generation by complex wavelet analysis
Publication Date: 2019.01.09 CORTEXICA VISION SYST
  • EP2353126B1 patent drawingFigure 1a~1b
  • EP2353126B1 patent drawingFigure 1c~2
  • EP2353126B1 patent drawingFigure 3

AI summary

The present invention relates to a method of generating a keypoint descriptor for identifying an object in an image or in a sequence of images, the keypoint descriptor being substantially invariant to a transformation of the object in the image. The method includes receiving object data representing an object for identification in an image; processing said object data to generate at least one basis function representing a feature having undergone at least one transformation or a transformation sequence across several consecutive frames, optionally using transformations that are out of the image plane to recognise objects from multiple views; modifying a prototype wavelet function based on said at least one basis function to generate a plurality of modified wavelet functions; comparing the plurality of modified wavelet functions with the at least one basis function; selecting a modified wavelet function of said plurality of modified wavelet functions based on the comparison of the plurality of modified wavelet functions with the at least one basis function; and processing an input image or input orientation field according to said selected modified wavelet function to generate the keypoint descriptor. The present invention further relates to a method of identifying an object in an image using a keypoint descriptor; and apparatus and computer software for implementing a method in accordance with the present invention.