Nonlinear correlation based on binary decompositions

Image decomposition is widely applied in many areas of image processing, for example subband decomposition. Examples of binary decomposition are threshold decomposition, bitmap decomposition, and the sliced orthogonal nonlinear generalized (SONG) decomposition. Our motivation for using binary decompositions is the possibility to optically implement those correlations with fast spatial light modulators that works in the binary regime.

Threshold decomposition is defined as
where
Next figure illustrates a threshold decomposition of a 0-255 gray-scale object.


A two-dimensional image with discrete gray levels can also be decomposed into a sum of disjoint elementary images eq(f(x,y)) defined as

The SONG decomposition of f(x,y) is
where


Correlation is frequently used of pattern recognition, however other nonlinear correlation based on binary decompositions are shown to have higher discrimination capabilities and noise robustness than common linear correlation.
Morphological correlation (MC) is defined as


where gq(x,y) and fq(x,y) are binary versions of g and f resulting from a threshold decomposition.

We defined the SONG correlation as the summation of multiple linear correlations between the binary slices of the two functions, g and f, weighting the correlation terms with a weight coefficient, Wij


At the origin, the autocorrelation sum for all the gray levels yields the total number of pixels in the objects. In early works on the SONG correlation we set Wij = ð ij, giving the same weight to all of the gray levels as

In this expression only the gray levels having the same values for the two images are correlated together after having their values set equal to one. The SONG correlation is very discriminating since it counts the number of points in the reference object and in the reference object and in the image that have the same gray levels at the same locations. To show the noise robustness of SONG correlation we used images that are degraded by substitutive noise. We have shown in our papers that SONG correlation is optimum in the maximum likelihood sense when the image is corrupted by a certain kind of substitutive noise.


Figure: (a) Input scene with the reference object below, and another object to be rejected. (b) Corrupted image with a high level of substitutive noise.


SEQUENTIAL JOINT TRANSFORM CORRELATOR

The nonlinear correlations that we are studying here are based on a sum of linear correlation between binary subimages of the targets and of the reference. The sums are performed on the joint power spectrum (JPS), which is the actual intensity output detected in the intermediate step of the JTC as

The setup is given by


The rotation and scale changes in the reference objects have been extensive tackled in pattern recognition. Some of the most effective methods for pattern recognition with in-plane rotations are based on the circular harmonic and the radial harmonic expansion for scale invariance.

We have defined the rotation invariant SONGC (RISONGC) as

where is the circular harmonic component (CHC) of the different binary slices of the reference object and are the polar coordinates.

We present some computer results to determine the performance of the method for rotation invariant pattern recognition when the input scene is corrupted with substitutive Gaussian noise






PAPERS ON THE SUBJECT: