References
107
where pi is a subimage of P with domain Xi E X. Notice that the inverse of this
statement is not true in general; however, we can use it as a test for the global
optimum of a registration function. In practice, the probability of false alarm
(the probability that an identified global optimum is just a local optimum) is
very low (Chen and Varshney 2001b), especially when the dimension of the
search space is high. Based on this idea, the global maximum of an MI based
registration function can be distinguished from local maxima and a robust
yet efficient global optimizer as the one shown in Fig. 3.15 can be successfully
designed.
3.7
Summary
Most of the existing similarity measures for intensity based image registration
problems depend on certain specific relationships between the intensities of
the images to be registered. For example, mean squared difference assumes
an identical relationship and cross correlation assumes a linear relationship.
If these assumptions are not strictly satisfied, registration accuracy may be
affected significantly. One example to illustrate this will be given in Chap. 7.
In this chapter, we have introduced the use of mutual information as a similarity measure for image registration. It does not assume specific relationship
between the intensity values of the images involved and hence, it is a very general similarity measure having been adopted for many different registration
applications involving different types of imaging sensors.
Tasks involved in MI based registration include joint histogram estimation
and global optimization of the MI similarity measure. These tasks have been
discussed in detail in this chapter. A phenomenon called interpolation-induced
artifacts was also discussed. A new joint histogram estimation scheme called
generalized partial volume joint histogram estimation was described. It eliminates or reduces the severity of this phenomenon. To build a robust yet efficient
global optimizer, a heuristic test to distinguish the global optimum from the
local ones was presented. Image registration examples using the techniques
introduced in this chapter will be provided in Chap. 7.
References
Chen H (2002) Mutual information based image registration with applications. PhD dissertation, Syracuse University
Chen H, Varshney PK (2000a) A pyramid approach for multimodality image registration
based on mutual information. Proceedings of 3rd international conference on information fusion 1, pp MoD3 9-15
Chen H, Varshney PK (2002) Registration of multimodal brain images: some experimental
results. Proceedings of SPIE Conference on Sensor Fusion: Architectures, Algorithms,
and Applications 6, Orlando, FL, 4731, pp 122-l33
Chen H, Varshney PK (2001a) Automatic two stage IR and MMW image registration algorithm for concealed weapon detection. lEE Proceedings, vision, image and signal
processing 148(4): 209-216
107
where pi is a subimage of P with domain Xi E X. Notice that the inverse of this
statement is not true in general; however, we can use it as a test for the global
optimum of a registration function. In practice, the probability of false alarm
(the probability that an identified global optimum is just a local optimum) is
very low (Chen and Varshney 2001b), especially when the dimension of the
search space is high. Based on this idea, the global maximum of an MI based
registration function can be distinguished from local maxima and a robust
yet efficient global optimizer as the one shown in Fig. 3.15 can be successfully
designed.
3.7
Summary
Most of the existing similarity measures for intensity based image registration
problems depend on certain specific relationships between the intensities of
the images to be registered. For example, mean squared difference assumes
an identical relationship and cross correlation assumes a linear relationship.
If these assumptions are not strictly satisfied, registration accuracy may be
affected significantly. One example to illustrate this will be given in Chap. 7.
In this chapter, we have introduced the use of mutual information as a similarity measure for image registration. It does not assume specific relationship
between the intensity values of the images involved and hence, it is a very general similarity measure having been adopted for many different registration
applications involving different types of imaging sensors.
Tasks involved in MI based registration include joint histogram estimation
and global optimization of the MI similarity measure. These tasks have been
discussed in detail in this chapter. A phenomenon called interpolation-induced
artifacts was also discussed. A new joint histogram estimation scheme called
generalized partial volume joint histogram estimation was described. It eliminates or reduces the severity of this phenomenon. To build a robust yet efficient
global optimizer, a heuristic test to distinguish the global optimum from the
local ones was presented. Image registration examples using the techniques
introduced in this chapter will be provided in Chap. 7.
References
Chen H (2002) Mutual information based image registration with applications. PhD dissertation, Syracuse University
Chen H, Varshney PK (2000a) A pyramid approach for multimodality image registration
based on mutual information. Proceedings of 3rd international conference on information fusion 1, pp MoD3 9-15
Chen H, Varshney PK (2002) Registration of multimodal brain images: some experimental
results. Proceedings of SPIE Conference on Sensor Fusion: Architectures, Algorithms,
and Applications 6, Orlando, FL, 4731, pp 122-l33
Chen H, Varshney PK (2001a) Automatic two stage IR and MMW image registration algorithm for concealed weapon detection. lEE Proceedings, vision, image and signal
processing 148(4): 209-216