Apply online to license this technology
Image registration geometrically aligns two images into one combined image that may provide more information to an observer than the single image. It’s a critical step in image analysis to identify changes in a scene. The technique is regularly used in weather forecasting, the creation of high-resolution images, integration of other sensor data into one image and in different forms of medical imaging. Image registration and fusion of multiple images into one is a well researched and commercialized data science. It becomes exponentially more difficult when trying to register and fuse time-varying image sources.
To deal with this issue, Air Force scientists have developed a unique method for registering and fusing time-varying image sources to provide the highest possible information rendering. The system aligns image sources by matching a target image to a reference image and minimizing visual registration error in a static sense. The system further selects target images which are best fused with a reference image using a dynamic, time-varying optimality maximum likelihood decision theory. This maximum likelihood decision theory is modified to account for time-variance using an orthogonal projection technique (optimal least squares solution) characterizing changing density functions.
The dynamic problem being solved is a natural extension of the static situation but involves many more issues. The Air Force solution has a number of new features to address this complexity. It employs two methods of optimization for the static registration and the dynamic fusion of image information. The initial calibration and registration of the multiple image sources are calibrated in an optimal sense that produces the lowest position error with a least squares measure using a method called singular value decomposition. A second optimality procedure is implemented to employ only those target images that add value in a sense to improve identification of objects. Elimination of those target images that only add correlation or bias is accomplished and this method eliminates those target images which add little new information by data mining.
- Target recognition from images which vary in time
- With this system it is not necessary to store a large number of reference images and thus the method saves considerable computation time and computer storage
- US patent 8,326,088 available for license