Object classification using image segmentation
View Patent ↗A method represents a class of objects by first acquiring a set of positive training images of the class of objects. A matrix A is constructed from the set of positive training images. Each row in the matrix A corresponds to a vector of intensities of pixels of one positive training image. Correlated intensities are grouped into a set of segments of a feature mask image. Each segment includes a set of pixels with correlated intensities. From each segment, a subset of representative pixels is selected. A set of features is assigned to each pixel in each subset of representative pixels of each segment of the feature mask image to represent the class of objects.
1. A method for representing a class of objects, comprising:
constructing a matrix A from a set of positive training images, in which each row in the matrix A corresponds to a vector of intensities of pixels at a particular pixel position in each one of the positive training images;
generating a segmented feature mask image, in which each segment of the segmented feature mask image corresponds to pixels with correlated intensities determined according to the vectors of intensities;
selecting, from each segment, a subset of representative pixels, in which a number of pixels in the subset is four or less; and
approximating each segment with a set of features, in which the set of features for the segment is determined from the corresponding subset of the representative pixels of each segment of the feature mask image to represent the class of objects.
2. The method of claim 1 , in which the set of features of each segment includes an approximate mean of the intensities and an approximate variance of the intensities of the segment.
3. The method of claim 1 , further comprising:
normalizing the set of positive training images before the constructing.
4. The method of claim 3 , in which each normalized image has about four hundred or fewer pixels.
5. The method of claim 1 , in which the columns in the matrix A correspond to leading components of a covariance matrix
C
=
1
N
AA
T
,
where N is the number of positive training images in the set, and T is a vector transform of the matrix A.
6. The method of claim 1 , in which the generating performs factor analysis.
7. The method of claim 6 , in which the factor analysis is K-means clustering.
8. The method of claim 2 in which μ i (x j ) is a true mean of the intensities of the set of pixels x in segment i of image j, and further comprising:
approximating the mean {circumflex over (μ)}(x j ) according to
μ
^
i
(
x
j
)
=
∑
j
=
1
k
x
j
k
,
where {x j } j=1 k is the subset of k representative pixels in the segment i of the image j.
9. The method of claim 8 , in which the selecting performs a greedy incremental search that minimizes
∑
j
=
1
n
(
μ
^
j
(
x
j
)
)
-
μ
i
(
x
j
)
)
2
,
for a next pixel of the set of pixels of the segment i to add to the subset of representative pixels.
10. The method of claim 9 , in which the approximate variance {circumflex over (σ)} i (x j ) of segment i of image j is
σ
^
i
(
x
j
)
=
∑
j
=
1
k
|
x
j
-
μ
i
(
x
j
)
|
.
11. The method of claim 1 , in which a total number of features of the segmented feature mask image is sixteen or less.
12. The method of claim 1 , in which a total number of segments is eight or less.
13. The method of claim 1 , further comprising:
searching exhaustively, for a set of selected classifiers, from a set of all available classifiers for the class of objects on single, pairs and triplets of the segments for a combination of segments and each available classifier that reject a largest number of non-object images using the positive training images and a set of negative training images.
14. The method of claim 13 , further comprising:
organizing the set of selected classifiers in a linear cascade.
15. The method of claim 14 , further comprising:
constructing a binary rejection table T, in which each column i represents a second set of negative training images, and each row j represents one of the selected classifier, and an entry T(i,j)=1 if a particular classifier i rejects a particular image j, and 0 if the particular image i is accepted.