Color Feature Detection
Color Feature Detection
CONTENTS
9.1 Introduction
........................................................................................................................203
9.2 Color Invariance
................................................................................................................205
9.2.1 Dichromatic Reflection Model
............................................................................. 205
9.2.2 Color
Invariants.....................................................................................................
206
9.2.3 Color Derivatives
.................................................................................................. 207
9.3 Combining Derivatives
..................................................................................................... 209
9.3.1 The Color Tensor
................................................................................................... 209
9.3.2 Color Tensor-Based Features
............................................................................... 210
9.3.2.1 Eigenvalue-Based Features
.................................................................. 210
9.3.2.2 Color Canny Edge Detection
............................................................... 211
9.3.2.3 Circular Object Detection
..................................................................... 212
9.4 Color Feature Detection: Fusion of Color Derivatives
................................................. 213
9.4.1 Problem
Formulation............................................................................................ 213
9.4.2 Feature Fusion
....................................................................................................... 214
9.4.3 Corner Detection
................................................................................................... 215
9.5 Color Feature Detection: Boosting Color
Saliency........................................................ 216
9.6 Color Feature Detection: Classification of Color
Structures........................................ 219
9.6.1 Combining Shape and Color
............................................................................... 219
9.6.2 Experimental Results
............................................................................................ 220
9.6.3 Detection of Highlights
........................................................................................ 221
9.6.4 Detection of Geometry/Shadow Edges
............................................................. 221
9.6.5 Detection of Corners
............................................................................................. 222
9.7 Conclusion
..........................................................................................................................
223
References.............................................................................................................
........................ 224
9.1
Introduction
The detection and classification of local structures (i.e., edges, corners, and
T-junctions) in color images is important for many applications, such as
image segmentation, image matching, object recognition, and visual tracking
in the fields of image processing and computer vision [1], [2], [3]. In general,
those local image structures are detected by differ- ential operators that are
commonly restricted to luminance information. However, most of the images
recorded today are in color. Therefore, in this chapter, the focus is on the use
of color information to detect and classify local image features.
203
Color Feature 20
Detection 4
The basic approach to compute color image derivatives is to calculate
separately the derivatives of the channels and add them to produce the final
color gradient. However, the derivatives of a color edge can be in opposing
directions for the separate color channels. Therefore, a summation of the
derivatives per channel will discard the correlation between color channels
[4]. As a solution to the opposing vector problem, DiZenzo [4] proposes the
color tensor, derived from the structure tensor, for the computation of the
color gradient. Adaptations of the tensor lead to a variety of local image
features, such as circle detectors and curvature estimation [5], [6], [7], [8]. In
this chapter, we study the methods and techniques to combine derivatives of
the different color channels to compute local image structures.
To better understand the formation of color images, the dichromatic
reflection model was introduced by Shafer [9]. The model describes how
photometric changes, such as shadows and specularities, influence the red,
green, blue (RGB) values in an image. On the basis of this model, algorithms
have been proposed that are invariant to different photometric phenomena
such as shadows, illumination, and specularities [10], [11], [12]. The extension
to differential photometric invariance was proposed by Geusebroek et al. [13].
Van de Weijer et al. [14] proposed photometric quasi-invariants that have
better noise and stability characteristics compared to existing photometric
invariants. Combining photometric quasi- invariants with derivative-based
feature detectors leads to features that can identify various physical causes
(e.g., shadow corners and object corners). In this chapter, the theory and
practice is reviewed to obtain color invariance such as shading/shadow and
illumination invariance incorporated into the color feature detectors.
Two important criteria for color feature detectors are repeatability, meaning
that they should be invariant (stable) under varying viewing conditions, such
as illumination, shad- ing, and highlights; and distinctiveness, meaning that
they should have high discriminative power. It was shown that there exists a
trade-off between color invariant models and their discriminative power [10].
For example, color constant derivatives were proposed [11] that are invariant
to all possible light sources, assuming a diagonal model for illumination
changes. However, such a strong assumption will significantly reduce the
discriminative power. For a particular computer vision task that assumes only
a few different light sources, color models should be selected that are
invariant (only) to these few light sources, result- ing in an augmentation of
the discriminative power of the algorithm. Therefore, in this chapter, we
outline an approach to the selection and weighting of color (invariant) models
for discriminatory and robust image feature detection.
Further, although color is important to express saliency [15], the explicit
incorporation of color distinctiveness into the design of salient points
detectors has been largely ignored. To this end, in this chapter, we review
how color distinctiveness can be explicitly incorporated in the design of image
feature detectors [16], [17]. The method is based upon the analysis of the
statistics of color derivatives. It will be shown that isosalient derivatives
generate ellipsoids in the color derivative histograms. This fact is exploited to
adapt derivatives in such a way that equal saliency implies equal impact on
the saliency map.
Classifying image features (e.g., edges, corners, and T-junctions) is useful for
a large num- ber of applications where corresponding feature types (e.g.,
material edges) from distinct images are selected for image matching, while
other accidental feature types (e.g. shadow and highlight edges) are
discounted. Therefore, in this chapter, a classification framework is discussed
Color Feature 20
Detection 5
to combine the local differential structure (i.e., geometrical information such
as edges, corners, and T-junctions) and color invariance (i.e., photometrical
information, such as shadows, shading, illumination, and highlights) in a
multidimensional feature space [18]. This feature space is used to yield proper
rule-based and training-based classifiers to label salient image structures on
the basis of their physical nature [19].
In summary, in this chapter, we will review methods and techniques solving
the following important issues in the field of color feature detection: to obtain
color invariance, such as
Color Feature 20
Detection 6
with shading and shadows, and illumination invariance; to combine
derivatives of the different color channels to compute local image structures,
such as edges, corners, circles, and so forth; to select and weight color
(invariant) models for discriminatory and robust image feature detection; to
improve color saliency to arrive at color distinctiveness (focus-of- attention);
and to classify the physical nature of image structures, such as shadow,
highlight, and material edges and corners.
This chapter is organized as follows. First, in Section 9.2, a brief review is
given on the various color models and their invariant properties based on the
dichromatic reflection model. Further, color derivatives are introduced. In
Section 9.3, color feature detection is proposed based on the color tensor.
Information on color feature detection and its appli- cation to color feature
learning, color boosting, and color feature classification is given in Sections
9.4, 9.5, and 9.6.
9.2 Color
Invariance
In this section, the dichromatic reflection model [9] is explained. The
dichromatic reflec- tion model explains the image formation process and,
therefore, models the photometric changes, such as shadows and
specularities. On the basis of this model, methods are dis- cussed containing
invariance. In Section 9.2.1, the dichromatic reflection model is intro- duced.
Then, in Sections 9.2.2 and 9.2.3, color invariants and color (invariant)
derivatives will be explained.
for C ∈ { R, G, B}, and where e (λ) is the incident light. Further, c b (λ) and
c s (λ) are the surface albedo and Fresnel reflectance, respectively. The
geometric terms mb and ms are the geometric dependencies on the body and
surface reflection component, λ is the wavelength, n is the surface patch
normal, s is the direction of the illumination source, and v is the direction of
the viewer. The first term in the equation is the diffuse reflection term. The
second term is the specular reflection term.
Let us assume that white illumination is when all wavelengths within the
visible spectrum have similar energy: e (λ) = e . Further assume that the
neutral interface reflection model holds, so that c s (λ) has a constant value
independent of the wavelength (c s (λ) = c s ). First, we construct a variable that
Color Feature 20
Detection 7
depends only on the sensors and the surface albedo:
With these assumptions, we have the following equation for the sensor
values from an object under white light [11]:
C w = emb (n, s)kC + ems (n, s, v)c s f (9.4)
with C w ∈ { Rw , G w , Bw }.
k
R
r ( R b , G b , Bb ) = (9.9)
kR + k G + k
B
g( Rb , G b , Bb ) = k G (9.10
kR + k G + k )
b( Rb , G b , Bb ) = B
kR + k
B (9.11
kG + k )
B
and hence, rgb is only dependent on the sensor characteristics and surface
albedo. Note that rgb is dependent on highlights (i.e., dependent on the
specular reflection term of Equation 9.4).
The same argument holds for the c 1 c 2 c 3 color space:
ma N
x
c 1 ( R b , G b , Bb ) = k k
arctan . k B}
R N
max{k
c 2 ( R b , G b , Bb ) = ,k
G
arctan
kG
(9.12) (9.13)
{ R B}
N
c 1 ( R b , G b , Bb ) = k (9.14
.k R
arctan )
B }
max{
Invariant properties for saturation
kG
min( R, G,
B)
S( R, G, B) = 1 (9.15)
R+G B
−
+
TA B L E9.1
Invariance for Different Color Spaces for Varying Image Properties
Syste Viewpoin Geometr Illumination Illumination Highligh
m t y Color Intensity ts
RG B — — — — —
rgb + + — + —
Hue + + — + +
S + + — + —
I — — — — —
c1 c2 c3 + + — + —
Note: A “+” means that the color space is not sensitive to the property; a “—”
means that it is.
are obtained by substituting the diffuse reflection term into the equation of
saturation:
min{k R , k G , k B }
S( Rb , G b , Bb ) = 1 — (9.16)
( R + k G +k B )
r = R2 + G2 + B 2
The scale factors follow from the Jacobian of the transformation. They ensure
that the norms of the derivative remains constant under the transformation,
hence |fx | = |f |. In the spherical coordinate system, the derivative vector
x is a
summation of a shadow–shading
variant part, Sx = (0, 0, r x ) T and a shadow–shading quasi-invariant part, x
given by Sc =
T
(r sin ϕθx , r ϕx , 0) [20].
The opponent color space is given
by
R —G
√
2 i
o1
R + i i
G —2 B
i i
o2 = √ (9.22
i )
6
o3
i
i
R+G+B
√
3
For this, the following transformation of the derivatives is obtained:
1
√ ( R x — G
o 1x x) i
2 i
o i 1 i
O (fx ) = f = = √ ( R + — 2 B )i (9.23)
o2 xG x i
x 6
x x
i
i
o3 x 1
√ ( Rx + G x + Bx )
3
The opponent color space decorrelates the derivative with respect to specular
x
changes.
T
The derivative is divided into a specular variant part,T
Ox = (0, 0, o
3x ) , and a specular quasi-invariant part Oc = (o 1x , o 2x , 0) .
The hue–saturation–intensity is given by
( N
o1
h arctan i
i o2 i
s = √ (9.24
i i )
o 12+ o 22
o3
The transformation of the spatial derivatives into the h si space decorrelates
the derivative with respect to specular, shadow, and shading variations,
( R ( BX — G X ) + G ( RX — BX ) + B (G X — RX ))
i
s hX 2( R2 + G 2 + B 2 — RG — RB — G B ) i
R (2 Ri
X — G X — BX ) + G (2G X — RX — BX ) + B (2 BX — RX
i — G X ) i
H (fX ) = fh = =
s j X
i
i
X
6( R2 + G 2 + B 2 — RG — RB — i
iX i
j GB )
( RX + G X +
BX ) (9.25)
√
3
The shadow–shading–specular variant is given by HX = (0, 0, i X ) T , and the
shadow–
shading–specular quasi-invariant is given Xby Hc = ( sh X , sX , 0) T .
Because the length of a vector is not changed by orthonormal coordinate
transformations,
the norm of the derivative remains the same in all three representations |fX | = |
fc | = |fo | = X X
h
|f X|. For both the opponent color space and the hue–saturation–intensity color
space, the
photometrically variant direction is given L 1 norm of the intensity. For the
by the spherical
coordinate system, the variant is equal to the L 2 norm of the intensity.
9.3 Combining
Derivatives
In the previous section, color (invariant) derivatives were discussed. The
question is how to combine these derivatives into a single outcome. A default
method to combine edges is to use equal weights for the different color
features. This naive approach is used by many feature detectors. For example,
to achieve color edge detection, intensity-based edge detec- tion techniques
are extended by taking the sum or Euclidean distance from the individual
gradient maps [21], [22]. However, the summation of the derivatives
computed for the dif- ferent color channels may result in the cancellation of
local image structures [4]. A more principled way is to sum the orientation
information (defined on [0, r )) of the channels
instead of adding the direction information (defined on [0, 2r )). Tensor
mathematics pro-
vide a convenient representation in which vectors in opposite directions will
reinforce one
another. Tensors describe the local orientation rather than the direction (i.e.,
the tensor of a vector and its 180◦ rotated counterpart vector are equal).
Therefore, tensors are convenient
for describing color derivative vectors and will be used as a basis for color
feature detection.
The Harris corner detector [24] is often used in the literature. In fact, the
color Harris operator H can easily be written as a function of the eigenvalues
of the structure tensor:
Further, the structure tensor of Equation 9.27 can also be seen as a local
projection of the derivative energy Ton two perpendicular axes [5], [7], [8],
namely, ul = ( l 0 ) T and i u2 = ( 0 l ) ,
(Gx, y ul ) · (Gx, y ul ) (Gx, y ul ) ·
Gul , u2 ‘ (9.33
(Gx, y u2 ) )
=
(Gx, y ul ) · (Gx, y u2 ) (Gx, y u2 ) ·
(Gx, y u2 )
in which Gx, y = ( gx g y ). From the Lie group of transformation, several other
choices of perpendicular projections can be derived [5], [7]. They include
feature extraction for circle, spiral, and star-like structures.
The star and circle detector is given as an example. It is based =√ ( x y )T ,
on ul l
x 2 +y2
which
coincides with the derivative pattern of a circular patterns = and
√ ( —y x ) T ,
u2 l
x 2 +y2 which
denotes the perpendicular vector field that coincides with the derivative
pattern of star-
like patterns. These vectors can be used to compute the adapted structure
tensor with
Equation 9.33. Only the elements on the diagonal have nonzero entries and are
equal to
x2 2xy y 2
gx ·gx g ·g g ·g 0 i
x2 + y2 + x 2 + x y x2 + y y i
H= 2 + i (9.34)
y y2 2xy 2
y
j
x2
0 g y ·g y 2 gx ·g y 2 gx ·gx
2
x +y 2 x + x +
— y2 + y2
where Al describes the amount of derivative energy contributing to circular
structures and
A2 the derivative energy that describes a star-like structure.
Curvature is another feature that can be derived from an adaption of the
structure tensor.
For vector data, the equation for the curvature is given by
To illustrate the performance, the results of Canny color edge detection for
several pho- tometric quasi-invariants is shown in Figure 9.la to Figure 9.le.
The image is recorded in three RGB colors with the aid of the SONY XC-003P
CCD color camera (three chips) and the
Color Feature 21
Detection 2
FIGURE9.1
(a) Input image with Canny edge detection based on, successively, (b) luminance derivative, (c)
RGB derivatives, (d) the shadow–shading quasi-invariant, and (e) the shadow–shading–specular
quasi-invariant.
Matrox Magic Color frame grabber. Two light sources of average daylight color
are used to illuminate the objects in the scene. The digitization was done in 8
bits per color. The results show that the luminance-based Canny (Figure 9.lb)
misses several edges that are correctly found by the RGB-based method
(Figure 9.lc). Also, the removal of spurious edges by pho- tometric invariance
is demonstrated. In Figure 9.ld, the edge detection is robust to shadow and
shading changes and only detects material and specular edges. In Figure 9.le,
only the material edges are depicted.
FIGURE9.2
(a) Detected circles based on luminance, (b) detected circles based on shadow–shading–specular
quasi-invariant, and (c) detected circles based on shadow–shading–specular quasi-invariant.
∂u ∂u i =l j =l i j
where cov(ui . u j ) denotes the covariance between two variables. From this
equation, it can be seen that if the function u is nonlinear, the resulting error,
σu , depends strongly on the values of the variables ul . u2 . · · . u N . Because
Equation 9.38 involves a linear combination of estimates, the error of the
combined estimate is only dependent on the covariances of the individual
estimates. So, through Equation 9.39, we established that the proposed
weighting scheme guarantees robustness, in contrast to possible, more
complex, combination schemes.
Now we are left with the problem of determining the weights wi in a
principled way. In the next section, we will propose such an algorithm.
or,
equivalently, V
N V
N V
2
σ i +
u
u = w2 σi 2
i =l j wi w j cov(ui . u j ) (9.4l)
•=
j =l
i
(9.43a)
i =l
—l <= wi <= l. i = l. · · . N
(9.43b) The constraint in Equations 9.43a ensures that all channels are fully
allocated, and the
constraint in Equation 9.43b limits the search
space for wi .
This model is quadratic with linear constraints and can be solved by linear
program-
ming [27]. When σu is varied parametrically, the solutions for this system will
result in mean–variance pairs representing different weightings of the feature
channels. The pairs
that maximize the expected u versus σu or minimize the σu versus expected
u, define the optimal frontier. They form a curve in the mean–variance plane,
and the corresponding weights are optimal.
(a) (b) (c) (d)
FIGURE9.3
(a) Lab image and (b) ground-truth for learning edges. Input image for the edge and corner
detection: on the left, the edge is indicated for the learning algorithm. (c) The χ -squared error of
the transformed image and the predicted expected value: here the edges have a very low
intensity. (d) The local signal-to-noise ratio for the transformed image. The edges have a higher
ratio.
A point of particular interest on this curve is the point that has the maximal
ratio between the expected combined output E (u) and the expected variance
2 u
σ . This point has the weights for which the combined feature space offers
the best trade-off between repeatability and distinctiveness.
In summary, the discussed selection model is used to arrive at a set of
weights to com- bine different color models into one feature. The expected
value of this feature E (u) is the weighted average of its component
expected values. The standard deviation of this combined feature will be less
than or equal to the weighted average of the component standard deviations.
When the component colors or features are not perfectly correlated, the
weighted average of the features will have a better variance-to-output ratio
than the individual components on their own. New features or colors can
always be safely added, and the ratio will never deteriorate, because zero
weights can be assigned to components that will not improve the ratio.
FIGURE9.4
(a) Results of the Harris corner detector. (b) Corners projected on the input image. The results of
the Harris corner detector, trained on the lower right cube.
all color spaces. The results of applying these weights to the same image are
shown in Figure 9.3c and Figure 9.3d. The edge is especially visible in the
signal-to-noise image. Using the weights learned on the edges with the Harris
operator, according to Equation 9.32, the corners of the green cube
particularly stand out (see Figure 9.4a and Figure 9.4b).
Another experiment is done on images taken from an outdoor object — a
traffic sign (see Figure 9.5a and Figure 9.5b). The weights were trained on one
image and tested on images of the same object while varying the viewpoint.
Again, the edges were defined by the first-order derivative in gauge
coordinates. The results of the Harris operator are shown in Figure 9.6. The
corner detector performs well even under varying viewpoints and illumination
changes. Note that the learning method results in an optimal balance between
repeatability and distinctiveness.
FIGURE9.5
The input image for the edge and corner detection: (a) the training image and (b) the trained
edges.
FIGURE9.6
Original images (top) and output (bottom) of the Harris corner detector trained on
red–blue edges.
the statistical properties of the image derivatives. The RGB color derivatives
are corre- lated. By transforming the RGB color coordinates to other systems,
photometric events in images can be ignored as discussed in Section 9.2,
where it was shown that the spatial derivatives are separated into
photometrical variant and invariant parts. For the purpose of color saliency,
the three different color spaces are evaluated — the spherical color space in
Equation 9.2l, the opponent color space in Equation 9.23, and the h si color
space in Equation 9.25. In these decorrelated color spaces, only the
photometric axes are influenced by these common photometric variations.
The statistics of color images are shown for the Corel database [28], which
consists of
40,000 images (black and white images were excluded). In Figure 9.7a to
Figure 9.7c, the
distributions (histograms) of the first-order derivatives, f x , are given for the
various color coordinate systems.
When the distributions of the transformed image derivatives are observed
from Figure 9.7, regular structures are generated by points of equal frequency
(i.e., isosalient surfaces). These surfaces are formed by connecting the points
in the histogram that occur the same number
100
–150
–150
Bx o3x rx
50
–75 –75
0
0 0
–50
75 75
FIGURE9.7
The histograms of the distribution of the transformed derivatives of the Corel image database in,
respectively, (a) the RGB coordinates, (b) the opponent coordinates, and (c) the spherical
coordinates. The three planes corre- spond with the isosalient surfaces that contain (from dark to
light), respectively, 90%, 99%, and 99.9% of the total number of pixels.
2l8 Color Image Processing: MetAods and
Applications
TA B L E9.2
The Diagonal Entries of I for the Corel Data Set
Computed for Gaussian Derivatives With σ = l
Parameter fX *fX *l s˜
fX ˜
S fo˜ ˜Xc
O fAX HcX
X X
c
( l 2 ( 2 2 ( 3
αA X + βA X γ AX = *IA(fX )*2 (9.44)
+ 2
( I I
X I A I
p X X
A(ff ) = p(f ) ÷ *IA(f )* = IX (9.45)
g (fX ) = IA (fX )
(9.46) By a rotation of the color axes followed by a rescaling of the axis, the
oriented isosalient
ellipsoids are transformed into spheres, and thus, vectors of equal saliency
are transformed
into vectors of equal
length.
Before color saliency boosting can be applied, the I-parameters have to be
initialized by
fitting ellipses to the histogram of the data set. The results for the various
transformations
are summarized in Table 9.2. The relation between the axes in the various
color spaces clearly confirms the dominance of the luminance axis in the RGB
2l8 Color Image Processing: MetAods and
Applications
cube, because I33 , the multiplication factor of the luminance axis, is much
smaller than the color-axes multiplica- tion factors, Ill and I22 . After color
saliency boosting, there is an increase in information context, see Reference
[l7] for more details.
To illustrate the performance of the color-boosting method, Figure 9.8a to
Figure 9.8d show the results before and after saliency boosting. Although
focus point detection is already an extension from luminance to color, black-
and-white transition still dominates the result. Only after boosting the color
saliency are the less interesting black-and-white structures in the image
ignored and most of the red Chinese signs are found.
(a)
(b) (c) (d)
(9.47)
K1 = f xx + f yy — ( f xx + f yy ) 2 xy
+4f2 (9.48)
xy
K2 = f xx + f yy + ( f xx + f yy )2
+4f2
FIGURE9.9
(a) Example image 1 (b) example image 2, and (c) example image 3. The images are recorded in
three RGB-colors with the aid of the SONY XC-003P CCD color camera (three chips) and the
Matrox Magic Color frame grabber. Two light sources of average daylight color are used to
illuminate the objects in the scene.
FIGURE9.10
(a) Test image, (b) Gaussian classifier, (c) mixture of Gaussians, and (d) k-nearest neighbor. Based
on the (training)
efficiency and accuracy of the results, the Gaussian or MoG are most appropriate for highlight
detection.
(a) (b) (c) (d)
FIGURE9.11
(a) Test image, (b) Gaussian classifier, (c) mixture of Gaussians, and (d) k-nearest neighbor. For
geometry–shadow detection, the best results are obtained by the Gaussian method. MoG and k-
nearest neighbor perform a bit less.
1
Gauss
0.9 mog
knn
0.8
0.7
0.6
Precision
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Recall
FIGURE9.12
Precision/recall graph for the classifiers of corners.
TA B L E9.3
Classifiers and Their Performance for
Corner Detection
Classifier Precisio Recall
n
Gaussian 34.9% 56.6%
Mixture of 77.8% 52.8%
Gaussians
k-Nearest neighbor 83.9% 58.5%
various settings have been examined. The results are shown in Figure 9.12.
The threshold providing the highest accuracy, and subsequently used in our
experiments, is 0.75.
From Table 9.3, it is observed that the k-nearest neighbor classifier provides
the highest performance. Examining the precision/recall graphs for the three
classifiers reveals that this method provides good performance. Further, the
MoG performs slightly better than the single Gaussian method.
The k-nearest neighbor classifier provides the best performance to detect
corners. Al- though the recall of all three methods is similar, the precision of
the k-NN classifier is higher.
9.7
Conclusion
In this chapter, we discussed methods and techniques in the field of color
feature detec- tion. In particular, the focus was on the following important
issues: color invariance, com- bining derivatives, fusion of color models, color
saliency boosting, and classifying image structures.
To this end, the dichromatic reflection model was outlined first. The
dichromatic reflec- tion model explains the RGB values variations due to the
image formation process. From the model, various color models are obtained
showing a certain amount of invariance. Then, color (invariant) derivatives
were discussed. These derivatives include quasi-invariants that have proper
noise and stability characteristics. To combine color derivatives into a single
outcome, the color tensor was used instead of taking the sum or Euclidean
distance. Tensors are convenient to describe color derivative vectors. Based
on the color tensor, var- ious image feature detection methods were
introduced to extract locale image structures such as edges, corners, and
circles. The experimental results of Canny color edge detec- tion for several
photometric quasi-invariants showed stable and accurate edge detection.
Further, a proper model was discussed to select and weight color (invariant)
models for dis- criminatory and robust image feature detection. The use of the
fusion model is important, as there are many color invariant models available.
In addition, we used color to express saliency. It was shown that after color
saliency boosting, (less interesting) black-and-white structures in the image
are ignored and more interesting color structures are detected. Finally, a
classification framework was outlined to detect and classify local image struc-
tures based on photometrical and geometrical information. High classification
accuracy is obtained by simple learning strategies.
In conclusion, this chapter provides a survey on methods solving important
issues in the field of color feature detection. We hope that these solutions on
low-level image feature detection will aid the continuing challenging task of
handling higher-level computer vision tasks, such as object recognition and
tracking.
References
[1] R. Haralick and L. Shapiro, Computer and Robot Vision, Vol. II, Addison-Wesley,
Reading, MA,
1992.
[2] C. Schmid, R. Mohr, and C. Bauckhage, Evaluation of interest point detectors, Int.
J. Comput.
Vision, 37, 2, 151–172, 2000.
[3] J. Shi and C. Tomasi, Good features to track, in Proceedings of IEEE Conference on
Computer Vision
and Pattern Recognition, Seattle, pp. 819–825, 1994.
[4] S. Di Zenzo, Note: A note on the gradient of a multi-image, Comput. Vision.
GrapAics.and Image
Process., 33, 1, 116–125, 1986.
[5] J. Bigun, Pattern recognition in images by symmetry and coordinate
transformations, Comput.
Vision and Image Understanding, 68, 3, 290–307, 1997.
[6] J. Bigun, G. Granlund, and J. Wiklund, Multidimensional orientation estimation with
applica-
tions to texture analysis and optical flow, IEEE Arans. on Patt. Anal. and MacAine
Intelligence, 13,
8, 775–790, 1991.
[7] O. Hansen and J. Bigun, Local symmetry modeling in multi-dimensional images,
Patt. Recog-
nition Lett., 13, 253–262, 1992.
[8] J. van de Weijer, L. van Vliet, P. Verbeek, and M. van Ginkel, Curvature estimation
in oriented
patterns using curvilinear models applied to gradient vector fields, IEEE Arans.
Patt. Anal. and
MacAine Intelligence, 23, 9, 1035–1042, 2001.
[9] S. Shafer, Using color to seperate reflection components, COLOR Res. Appl., 10,
210–218, Winter
1985.
[10] T. Gevers and H. Stokman, Robust histogram construction from color
invariants for ob-
ject recognition, IEEE Arans. on Patt. Anal. and MacAine Intelligence (PAMI), 26,
1, 113–118,
2004.
[11] T. Gevers and A. W. M. Smeulders, Color based object recognition, Patt.
Recognition, 32, 453–464,
March 1999.
[12] G. Klinker and S. Shafer, A physical approach to color image understanding, Int.
J. Comput.
Vision, 4, 7–38, 1990.
[13] J. Geusebroek, R. van den Boomgaard, A. Smeulders, and H. Geerts, Color
invariance, IEEE
Arans. Patt. Anal. MacAine Intelligence, 23, 12, 1338–1350, 2001.
[14] J. van de Weijer, T. Gevers, and J. Geusebroek, Edge and corner detection by
photometric quasi-
invariants, IEEE Arans. Patt. Anal. and MacAine Intelligence, 27, 4, 625–630, 2005.
[15] L. Itti, C. Koch, and E. Niebur, Computation modeling of visual attention, Nat. Rev.
Neuroscience,
2, 194–203, March 2001, 365–372.
[16] J. van de Weijer and T. Gevers, Boosting color saliency in image feature detection,
in International
Conference on Computer Vision and Pattern Recognition, San Diego, CA, USA, pp.
365–372, 2005.
[17] J. van de Weijer, T. Gevers, and A. Bagdanov, Boosting color saliency in image
feature detection,
IEEE Arans. Patt. Anal. and MacAine Intelligence, 28, 1, 150–156, 2006.
[18] T. Gevers and H. Stokman, Classification of color edges in video into shadow-
geometry, high-
light, or material transitions, IEEE Arans. on Multimedia, 5, 2, 237–243, 2003.
[19] T. Gevers and F. Aldershoff, Color feature detection and classification by learning,
in Proceedings
IEEE International Conference on Image Processing (ICIP), Vol. II, pp. 714–717,
2005.
[20] J. van de Weijer, T. Gevers, and A. Smeulders, Robust photometric invariant
features from the
color tensor, IEEE Arans. Image Process., 15, 1, 118–127, 2006.
[21] S.D. Zenzo, A note on the gradient of a multi-image, Comput. Vision. GrapAics.
and Image Process.,
33, 116–125, 1986.
[22] G. Sapiro and D.L. Ringach, Anisotropic diffusion of multivalued images
with app-
lications to color filtering, IEEE Arans. Patt. Anal. and MacAine Intelligence, 5,
1582–1586,
1996.
[23] G. Sapiro and D. Ringach, Anisotropic diffusion of multivalued images with
applications to
color filtering, IEEE Arans. Image Process., 5, 1582–1586, October 1996.
[24] C. Harris and M. Stephens, A combined corner and Manchester, UK edge detector,
in Proceedings of tAe FourtA Alvey Vision Conference, Manchester, UK, Vol. 15,
1988, pp. 147–151.
[25] D.H. Ballard, Generalizing the Hough transform to detect arbitrary shapes, Patt.
Recognition,
12, 111–122, 1981.
[26] H. Markowitz, Portfolio selection, J. Finance, 7, 1952.
[27] P. Wolfe, The simplex method for quadratic programming, Econometrica, 27,
1959.
[28] C. Gallery, www.corel.com.