US 5,848,193 AGranted

Wavelet projection transform features applied to real time pattern recognition

AssigneeThe United States of America as represented by the Secretary of the Navy

InventorJoseph P. Garcia

Life of this patent

Estimating the term…

Abstract

A computer vision system having improved means for characterizing the parameters of edge segments of an unknown object is disclosed. The unknown object is transformed by a digital camera into an image and transformed by a wavelet projection transform so as to decompose the unknown objects into their constituent multi-scale edge segment features and provide recognition of the unknown object by relating their constituent edge features to each other.

Claims11

Dependent claims

A method of operating a computer vision system having a digital camera outputting a digital image, for recognizing an unknown object having multi-scale edge segment components by converting the digital image into a particular representation, said method comprising the steps of:

(a) transforming by a wavelet projection transform (WPT), said digital image into a first signal having at least four (4) dimensions and having corresponding space dimensions and defining a wavelet projection transform (WPT) domain;
(b) processing said four dimensional transform to find a set of representative quantities that correspond to constituent multi-scale edge segment (MSES) features of the digital image so as to yield an efficient edge feature decomposition of said digital image;
(c) grouping at least two MSES representative quantities into groups whereby a particular group is referable to as a location of a particular MSES representative quantity;
(d) deriving a set of invariant relations having associated parameters between the at least two MSES representative quantities of said group of step (c), said derived set being in said four dimensional transform space and having specific invariances which include shift, scale and rotation invariance within an aspect view;
(e) applying steps (a), (b), (c), and (d) to multiple aspect views of a prototype object and deriving a set of invariants having associated parameters for the said multiple aspect views and forming a library of said invariants for said prototype objects, and applying said steps (a), (b), (c) and (d) to the image of an unknown object;
(f) correlating said set of invariant relations of said step (d) to said set of invariant relations of step (e) so as to generate at least four dimensional digital signals which are representative of a particular class of prototype objects and have multidimensional phases which are functions of corresponding MSES representative quantities derived from the unknown image and parameters associated with the invariants derived from the prototype object;
and
(g) superposing the said representative signals of step (f) and storing the results such that the correlations and matches of digital image MSES components corresponding to those of the library of prototype objects result in coherent superpositioning of said representative signals and then determining the degree of coherence by performing a digital matched filtering operation with a reference signal derived from a representative signal of a particular class of prototype objects such that a correlation peak indicates the recognition of the unknown object in the digital image and the position of the correlation peak indicates the position of the recognized unknown object.

5 more elements

The method according to claim 1, wherein said digital image is represented by a two dimensional field of intensities and consists of an ordered list of digital values that may be formed into rows and columns of pixels and has two canonical axes corresponding to the directions along said rows and columns, said digital image being transformed by a wavelet projection transform (WPT) according to the following steps:

(a) performing a Cartesian rotational transform of said digital image for a predetermined number of times to incrementally varying orientations between 0 and 180 degrees;
(b) generating a copy of said digital image for each of said orientations;
(c) integrating about pixels of said rotated digital image copies of step (b) of claim 6 for each instantaneous orientation along a canonical axis of pixels of said digital image copy by using a digital low pass filter applied recursively to its downsampled output thus generating multiple levels of output, said digital low pass filter having filter coefficients selected from a low pass section of a discrete wavelet transform filter bank, said integrating being separately performed on each of said digital image copies of step (b) of claim 6;
and
(d) differentiating said digital image copies of step (c) of claim 6 for each orientation along the transverse direction relative to said canonical axis of step (b) of claim 6 by using the continuous wavelet transform (CWT), said differentiating being separately performed on said rotated images having a range of scales, with instantaneous scale parameters occurring as a function of the level of the said digital low pass filter of step (c) of claim 6, whereby each scale of the CWT corresponds uniquely to a level on the said digital low pass filter.

2 more elements

Dependencies are inferred from the claim text.

patentgrep

Loading patent…

Abstract

(a) transforming by a wavelet projection transform (WPT), said digital image into a first signal having at least four (4) dimensions and having corresponding space dimensions and defining a wavelet projection transform (WPT) domain;
(b) processing said four dimensional transform to find a set of representative quantities that correspond to constituent multi-scale edge segment (MSES) features of the digital image so as to yield an efficient edge feature decomposition of said digital image;
(c) grouping at least two MSES representative quantities into groups whereby a particular group is referable to as a location of a particular MSES representative quantity;
(d) deriving a set of invariant relations having associated parameters between the at least two MSES representative quantities of said group of step (c), said derived set being in said four dimensional transform space and having specific invariances which include shift, scale and rotation invariance within an aspect view;
(e) applying steps (a), (b), (c), and (d) to multiple aspect views of a prototype object and deriving a set of invariants having associated parameters for the said multiple aspect views and forming a library of said invariants for said prototype objects, and applying said steps (a), (b), (c) and (d) to the image of an unknown object;
(f) correlating said set of invariant relations of said step (d) to said set of invariant relations of step (e) so as to generate at least four dimensional digital signals which are representative of a particular class of prototype objects and have multidimensional phases which are functions of corresponding MSES representative quantities derived from the unknown image and parameters associated with the invariants derived from the prototype object;
and
(g) superposing the said representative signals of step (f) and storing the results such that the correlations and matches of digital image MSES components corresponding to those of the library of prototype objects result in coherent superpositioning of said representative signals and then determining the degree of coherence by performing a digital matched filtering operation with a reference signal derived from a representative signal of a particular class of prototype objects such that a correlation peak indicates the recognition of the unknown object in the digital image and the position of the correlation peak indicates the position of the recognized unknown object.

5 more elements