A multimodal image registration technique for structured polygonal scenes

A technique for multimodal image registration based on a hypothesize-and-test approach is presented. The technique is based on aligning edge elements from the two input images, since they often originate from physical discontinuities which are likely to be detected by both sensors. The design has been specifically adapted for robust operation on images of regular objects with few distinct structural axes, in the context of automated inspection. The hypotheses are formed by speculating a correspondence between the pairs of parallelograms from the two input images, and evaluated according to the quality of the match of a transformed edge image. The best hypothesis is finally refined by a nonlinear optimization algorithm. The technique has been tested on 42 pairs of 14 infrared and 3 RGB images of an unevenly heated metal prism, and the obtained results are reported.