Image registration transforms different sets of data into a single coordinate system and is thus critical for comparison and integration of data obtained from different measurements. This technique is useful in the medical imaging field as well as in other monitoring, observation and tracking applications. Image registration is complex especially when the images to be correlated exhibit intensity differences and spatial variations. One of the main issues arises when intensity differences are due to the sensor itself i.e., in registration with different types of imaging sensors, or registration of pf speckle-limited images. A number of “enthropic” methods (e.g., level set matching, Jensen Difference minimization, mutual information maximization) have been developed to overcome this problem. However, they are largely based on density estimation techniques, which are exceedingly difficult as feature dimension become high. Thus enthropic registration methods have been limited to low dimensional feature spaces.
Researchers at the University of Michigan have developed a novel algorithm that provides extensions to enthropic similarity measures thus breaking the computation bottleneck for high dimensional features. The guiding principle behind these extensions is the use of continuous quasi-additive power weighted graphs, such as the minimal spanning tree and k-Nearest Neighbor graph, to estimate enthropic similarity measures. Different graph length functionals will allow the approximation of a wide variety of enthropic matching criteria without the need to explicitly estimate densities or histograms. Advantages of developed approach have been demonstrated for speckle limited ultrasound breast imaging, computing tumor size/volume changes in response to therapy, geo-registration, and face retrieval.
Applications and Advantages
Applications • Medical imaging • Geo-registration (i.e. satellite tracking, cartography updating) • Face retrieval • Computer vision Advantages • Drastically reduces computing time for enthropic image registration techniques, which allows for use of these methods at high feature dimension applications.