US 8,620,082 B1 · Sonar image texture segmentation

Abstract

Systems and methods for sonar image feature extraction, which fit a parameterized statistical model to sonar image data to extract features that describe image textures for follow-on pattern classification tasks. The systems and methods estimate the parameters of an assumed statistical model from the pixel values in the sonar images. The parameters of the distribution can then be used as features in a discriminant classifier to perform tasks such as texture segmentation or environmental characterization. The systems and methods divide a sonar image into separate overlapping cells. The ACF parameters of each cell are sequentially estimated using various methods. Based on the ACF parameters obtained, each cell is assigned a label via a pattern classification algorithm. The labeled image can be used in target recognition, environmental characterization and texture segmentation.

Claims11

Dependent claims

A method of texture segmentation for a sonar image, S(x,y), comprising using a processor to perform the steps of:

obtaining initial estimates of a first set of parameters of said sonar image;
obtaining an estimate of an intensity autocorrelation function (ACF) of said image;
initializing a second set of parameters, Θi={αi, [μXiμYi], Σi} of said sonar image based on said first set of parameters;
initializing a likelihood function based on said first and second sets of parameters;
maximizing said likelihood function to obtain updated parameters;
calculating an updated intensity ACF based on said updated parameters;
and
iteratively returning to maximizing said likelihood function until convergence of said updated intensity ACF.

5 more elements

A method of segmenting textures in a sonar image, S(x,y), comprising using a processor to perform the steps of:

dividing said image into a plurality of (N×O) cells, wherein said cells overlap;
obtaining initial estimates of a first set of parameters of a first of said cells;
obtaining an estimate of an intensity autocorrelation function (ACF), {circumflex over (R)}l(X, Y), of said first cell via discrete two-dimensional Fourier transform:
R
^
l
⁡
(
X
,
Y
)
=
1
NO
·
F
-
1
⁢
{
F
⁢
{

S
⁡
(
x
,
y
)

.
2
*
⁢
conj
(
F
⁢
{

S
⁡
(
x
,
y
)
⁢
❘
2
}
)
}
;
initializing a second set of parameters, Θi={αi, [μXiμYi], Σi}, of said first cell based on said first set of parameters;
initializing a likelihood function based on said first and second sets of parameters;
maximizing said likelihood function to obtain updated parameters;
calculating an updated intensity ACF based on said updated parameters;
iteratively returning to maximizing said likelihood function until convergence of said updated intensity ACF;
storing cell parameters associated with said updated intensity ACF;
iteratively returning to obtaining initial estimates of a first set of parameters for each of the plurality of cells;
and
segmenting textures of said sonar image based on said cell parameters stored for said plurality of cells.

63 more elements

Dependencies are inferred from the claim text.

Sonar image texture segmentation

Life of this patent

Sonar image texture segmentation

Life of this patent

Abstract

Claims11