I have sets of monochrome images of a manuscript leaf that I'd like to register using a full-up warping approach based on extracted features. These images were recorded at different wavelengths and at different times. In consequence, the objects are slightly and nonuniformly warped (due to temperature and humidity change during image sequence capture) and slightly different in focal quality (even with a good apochromatic lens).
I wonder about the influence of the boundary of the leaf (the ROI) on registration for two different cases:
(i) the leaf is placed on a background that appears in the images as a margin with well defined edges and has fairly uniform content much different in brightness and variance than the leaf itself (the ROI);
(ii) a mask object (a so-called "window matte") is placed over the leaf during image capture so that each image shows a rectangular region of the leaf surrounded by a dark margin. Within the opening of the mask, the leaf may be translated, rotated, scaled, warped, etc., but its boundary is always the same rectangular shape. The images could be cropped to this rectangular window, but the pixels on the edge of the ROI (which would then be the entire cropped image) would still correspond to different points of the leaf.
In both cases my concern is the influence of the ROI boundary and its exterior on the registration process if the images are not cropped to eliminate non-leaf margin.
I wonder if the high contrast at the boundary of the leaf ROI, and the uniformity of the margin around it, causes ROI boundary or exterior pixels or features to be overweighted in approaches (such as SIFT) that use geometric features. Or, just for case (i), since the boundary consists of the same set of points of the leaf, and only the contour shape of the boundary changes, does the boundary provide an exceptionally valuable set of features for registration?
Is there an effective way to reduce any inappropriately large weighting of the boundary if it occurs? Is there a way in the BUnwarpJ or Extract SIFT Correspondences plugins to have the registration calculation exclude features or other content outside of a defined ROI? Alternatively, would blurring or some other apodization of the boundary zone (of some user-defined width) be a sensible way to influence the choice or weighting of features that are extracted?
Can anyone recommend particular ways of minimizing this boundary effect if it occurs? Is there discussion of this in texts on registration methods?
Thanks.
Bill Christens-Barry
--
ImageJ mailing list:
http://imagej.nih.gov/ij/list.html
Locked
