DualAlign has developed sophisticated computer vision technology, i2Align,
that can automatically stitch, organize, align and view images even
when the images are taken from different cameras, at different times,
under different conditions. One of the fundamental building blocks to
i2Align is DualAlign’s image registration and recognition
technology.
Image registration and recognition is the
process that identifies features, objects, or elements shared between
images taken at the same time or later, and with the same sensor or a
different one. Image registration and recognition is an important
building block for imaging applications, because it is necessary in
order to be able display and analyze image data within these
applications. At best, without image registration and recognition,
users would be forced to view images side by side or attempt to
manually overlay images making many image analysis applications in
markets like medical; satellite, thermal; and photographic imaging
impractical or too slow for mass consumption.
DualAlign’s
image registration and recognition technology is considered the most
advanced technology of its kind providing a level of automation and
robustness previously thought to be unobtainable. DualAlign’s
image registration and recognition technology has the ability to
automatically process images taken using different imaging modalities,
having low overlap in their fields of view, taken when there are
physical and illumination changes in the scene between images, and
showing large changes in image scale. While other technologies
exist for image recognition, none come near the capabilities of
DualAlign’s image registration and recognition technology
for full automation and applicability in a wide-variety of applications.
Dual-Align’s
image registration and recognition typically involves the following
steps: Recognition, Registration, and Decision.
Recognition
During Recognition (initialization) a set of candidate matches between the two images is generated. Each candidate match includes (a) an image keypoint pair, (b) image regions ("bootstrap regions") around the pair and (c) a transformation function. This transformation function will align the two images, but only within the bootstrap region. An example of a pair images, one taken in the winter and one in the summer, along with a keypoint pair and the corresponding bootstrap regions is displayed in Figure 1. In essence, each candidate match is an "educated guess" that the two associated regions from the regions are the same. Some of these matches are correct, but many of them may be wrong.
Figure 1
Registration
The Registration Algorithm attempts to align the images starting from one of the candidate matches generated during initialization. This sophisticated algorithm is capable of ignoring differences in structure and illumination between images, automatically determining what is consistent between the images in order to generate a transformation function. It does this by gradually “growing” the bootstrap region and refining the transformation function to eventually apply to all of both images. The vast majority of correct initial keypoint matches are grown into correct final alignments between the images.
Decision
A very sophisticated Decision Criteria Algorithm then determines if the alignment is acceptable. If so, the transformation function used to generate the alignment is taken as correct and the whole three-step procedure ends. If not, the registration algorithm is applied to the next candidate match. Experiments have shown that the Decision Criteria Algorithm makes the correct decision nearly 100% of the time. In addition, if the overall three-step algorithm is given two images that do not depict the same thing, the Decision Criteria will reject all initial keypoint matches and refined results, indicating that the two images can not be registered.
An example of two correctly aligned images is displayed in Figure 2.
Figure 2
