Professional Documents
Culture Documents
I. INTRODUCTION
Manuscript received March 27, 1997; revised April 3, 2001. This work was
supported in part by NIH Grant GM07717, NASA Grant NAG10-0155, Texas
Advanced Research Program Grant ARP002-1996, and Canada NSERC Grant
RGPIN239031. The associate editor coordinating the review of this manuscript
and approving it for publication was Dr. Dapang Chen.
X.-P. Zhang is with the Department of Electrical and Computer Engineering,
Ryerson Polytechnic University, Toronto, ON, M5B 2K3 Canada (e-mail:
xpzhang@ieee.org).
M. Desai is with the Division of Engineering, University of Texas, San Antonio, TX 78249 USA.
Publisher Item Identifier S 1057-7149(01)05435-5.
1021
(4)
Assume
where
PDF of the class ;
a priori probability of class
in image ;
PDF of the class ;
a priori probability of class
in image .
The above classification problem is solved by the standard
Bayes classifier
(6)
where
and
Fig. 2. Dashed line indicates the PDF p (x). Two solid lines indicate
P (C )p (x) and P (C )p (x), respectively. The Bayes threshold and the
proposed threshold are indicated.
Assuming that
and
have one point of intersection
, as illustrated in Fig. 2, the above classifier is equivalent to the
following threshold detection criterion:
(7)
where
is chosen so that
(8)
(9)
There are two main problems in using the above Bayes classifier for segmentation of targets in an image. First of all, images may have a rather complex structure; they contain other
classes. For exampl, in a mammogram, the object of segmentation is the tumor which generally has a higher greyscale value
than the remainder of the tissue. Within a mammogram, there
are calcifications, dense tissue, and various kinds of noise, etc.,
in addition to the normal tissues and possible tumors (targets).
The normal tissues and possible targets also contain some singular pixels. Hence, many pixels may be misclassified by using
simple threshold detection method on original mammograms.
Second, in practice, we do not actually know the PDF of any
class as well as the a priori probability of every class. In addition, the probability distributions often vary greatly in different
images so that a general distribution cannot be obtained to describe all images.
By analyzing the image at a coarser scale, singularities can be
removed. However, the PDF of any class as well as the a priori
probability are still unknown in the low-resolution images. It is
difficult to choose a suitable threshold automatically. Of course,
trial and error is a possible approach. However, it is inefficient
and difficult to use in practice. In the following, a method based
1022
and
if
and
(12)
otherwise
1023
where
denotes a pixel of transformed image, the
matrix
denotes a matrix of input pixels,
denotes the
denotes the horizontal
vertical wavelet transform vector, and
wavelet transform vector.
Normally, the input pixels in a neighborhood should belong
to the same class and can be treated as independent identically
distributed (i.i.d.) random variables with the same distribution.
Then according to the central-limit theorem [16], as increases,
approaches a Gaussian distribution.
the distribution of
Therefore, we can approximately treat the gray level of each
class in the transformed image as having a Gaussian distribution.
Hence, the performance of the presented method will be analyzed based on the Gaussian distribution. Assume both classes
in transformed image are Gaussian distributed, i.e.,
(15)
and
(16)
Then, for a Bayes threshold , we have
Fig. 3.
is the set of
(17)
i.e.,
and
(18)
(13)
is selected as the
without loss of
For simplification, first assume
and
.
generality. Then take
Normalize with respect to class , i.e., let
and denote
. Equation (18) can be
rewritten as
(19)
(14)
It is also possible to treat the segmented image as a new original image, which may include more detailed classes. Then, a
wavelet analysis of the PDF can be used again to classify the
different classes within the segmented area. The process can be
iterated again until certain criteria are satisfied.
and
. From the above
where
, , and
equation, it can be seen that is only of concern to
.
for the above equation in
Obviously, only the solution of
should be taken. The solutions which are out
the range
are meaningless for the classification of two
of the range
classes.
1024
Assume
and
class
and
are the a priori probabilities of
in image . Then (5) can be rewritten as
(20)
and
where
, we have
For the proposed threshold
(21)
and
(22)
Substituting (15) and (16) into (21), we get
Dashed line indicates the PDF p (x). Two solid lines indicate
and P (C )p (x), respectively. The Bayes threshold is
indicated.
Fig. 4.
P (C )p (x)
(23)
We take the terms with order less than three. Substituting (15)
and (16) into the above equation, and using (18) and (21), we
obtain
(24)
(29)
(25)
and the probability of misclassification error of class
(30)
where
is
It can also be obtained from (22) that
(26)
is defined as [17]
(27)
and
(28)
(31)
must be large enough to choose
Equation (31) shows that
and
the local minimum point as the threshold, i.e.,
. In fact, if (31) is not satisfied, the two
classes will have a large overlap so that the misclassification
error will be large even using the Bayes threshold. Fig. 4 shows
an example.
becomes
From these equations, it is easy to see that when
becomes larger and then
will become smaller.
larger,
and
have no simple and monotonic relationHowever,
. In the following figures, the performance
ship with and
of the threshold in different cases is given.
Because of the symmetry of the Gaussian function, we can
without loss of generality.
assume
In Fig. 5, the curves show how the threshold and its perforbut constant
, where
mance change with different and
ranges in
,
and
.
1025
()
Fig. 6. Dashed line indicates the PDF p x . Two solid lines indicate
P C p x and P C p x , respectively. We have a relatively larger
P for the Bayes threshold .
( ) ()
()
( ) ()
=5
=2
= 1 2 ... 7
1
()
Fig. 7. Dashed line indicates the PDF p x . Two solid lines indicate
P C p x and P C p x , respectively. When is very small,
also becomes very small. Note that the dashed line and two solid lines almost
overlap completely.
( ) ()
( ) ()
1026
=5
= 2 01
= 1 2 ... 9
1
10242 1024
Fig. 10. Histograms (PDFs) of the four transformed images in Fig. 9(b)(e).
The top one shows the histogram of the original image.
1027
Fig. 11. Five-scale wavelet transform for the histogram of the image in scale
2 [Fig. 9(c)].
Fig. 12. (a) Segmented areas using . (b) The masked image using the
segmented image I
in (a) by I
1 I
I.
Fig. 13. (a) Segmented areas using . (b) Masked image using the segmented
image I
in (a) by I
1 I
I.
1028
Fig. 14. (a) Segmented areas using . (b) Masked image using the segmented
image I
in (a) by I
1 I
I.
Fig. 16. PDFs of the four transformed images in Fig. 15(b)(e). The top one
shows the histogram of the original image.
256 256
1029
Fig. 20. (a) Segmented areas using . (b) Masked image using the segmented
image I
in (a) by I
1 I
I.
V. CONCLUSION
Fig. 17. Five-scale wavelet transform for the histogram of the image in scale
2 [Fig. 15(c)].
Fig. 18. (a) Segmented areas using . (b) Masked image using the segmented
image I
in (a) by I
1 I
I.
We have presented a new general method for the segmentation of bright targets in an image. Wavelet transforms are used
in the new method and the Bayes classifier is employed for the
segmentation problem. An approach for choosing the threshold
adaptively by looking for the global local minima of the PDFs
of wavelet transformed images is proposed. Based on the assumption of Gaussian distributions, the adaptive threshold by
the new method is compared with the Bayes threshold. It is
shown that in general practical cases, the performance of the
proposed threshold is often very close to the Bayes threshold,
which is the optimal threshold from the statistical point of view.
The examples show that the new method is effective to segment the tumors in mammograms and it can also be used in other
segmentation applications.
To select an appropriate scale of image is an important step of
this method. The general principles for choosing scaling channel
along with the wavelet are presented in this paper. Because the
method works under the concept of general multiscale analysis,
we can expect the choices are image-dependent, however, not
very image-sensitive as evident by the fact that the same parameters were used for both mammograms and FLIR images.
After the processing using the new method, other algorithms,
such as boundary based approaches can also be combined with
the new method depending on what is of interest in various applications.
Furthermore, although the new method is based on the distribution of the greylevel of image, it is also possible to use the
method to other specific features. That is, a new image can be
constructed by extracting some features from the original image
for segmentation. For example, we have used it to detect the singularities, such as the microcalcifications in mammograms, and
it also shows good performance [20].
ACKNOWLEDGMENT
Fig. 19. (a) Segmented areas using . (b) Masked image using the segmented
image I
in (a) by I
1 I
I.
1030
[2] S. A. Dudani, B. Smithgall, and P. Robins, Target acquisition and extraction from cluttered backgrounds, in Proc. SPIE, IR Technol. Target
Detection Classification, vol. 302, 1981, pp. 4350.
[3] D. T. Long, I. E. Abdou, and S. Sutha, Parallel algorithm for modeldirected real-time image segmentation, in Proc. SPIE, Realtime Image
Processing, vol. 302, 1990, pp. 27.
[4] S. M. Dunn, D. Harwood, and L. S. Davis, Local estimation of the
uniform error threshold, IEEE Trans. Pattern Anal. Machine Intell., vol.
PAMI-6, no. 6, pp. 742747, 1984.
[5] A. K. Jain, Fundamentals of Digital Image Processing. Englewood
Cliffs, NJ: Prentice-Hall, 1986.
[6] P. K. Sahoo, S. Soltani, and A. K. C. Wong, A survey of thresholding
techniques, Comput. Vis. Graph. Image Process., vol. 41, pp. 233260,
1988.
[7] J. S. Weszka and A. Rosenfeld, Threshold evaluation techniques,
IEEE Trans. Syst., Man, Cybern., vol. SMC-8, no. 8, pp. 622629, 1978.
[8] B. Chanda and D. D. Majumder, A note on the use of the gray-level
co-occurrence matrix in threshold selection, Signal Process., vol. 15,
pp. 149167, 1988.
[9] W.-N. Lie, Automatic target segmentation by locally adaptive image
thresholding, IEEE Trans. Image Processing, vol. 41, pp. 10361041,
July 1995.
[10] S. Mallat, A theory for multiresolution signal decomposition: The
wavelet representation, IEEE Trans. Pattern Anal. Machine Intell.,
vol. 11, no. 7, pp. 674693, July 1989.
[11] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: SIAM,
1992.
[12] O. Rioul, A discrete-time multiresolution theory, IEEE Trans. Signal
Processing, vol. 41, pp. 26912606, Aug. 1993.
[13] O. Rioul and P. Duhamel, Fast algorithms for discrete and continuous
wavelet transforms, IEEE Trans. Inform. Theory, vol. 38, pp. 569586,
Mar. 1992.
[14] M. J. Shensa, The discrete wavelet transform: Wedding the trous
and mallat algorithms, IEEE Trans. Signal Processing, vol. 40, pp.
24642482, Oct. 1992.
[15] A. N. Akansu and R. A. Haddad, Multiresolution Signal Decomposition. New York: Academic, 1992.
[16] A. Papoulis, Probability, Random Variables, and Stochastic Processes. New York: McGraw-Hill, 1965.
[17] H. L. Van Trees, Detection, Estimation, and Modulation Theory. Part
I. New York: Wiley, 1968.