Fast phase-based registration of multimodal image data

An interesting problem in pattern recognition is that of image registration, which plays an important role in many vision-based recognition and motion analysis applications. Of particular interest among registration problems are multimodal registration problems, where the images exist in different feature spaces. State-of-the-art phased-based approaches to multimodal image registration methods have provided good accuracy but have high computational cost. This paper presents a fast phase-based approach to registering multimodal images for the purpose of initial coarse-grained registration. This is accomplished by simultaneously performing both globally exhaustive dynamic phase sub-cloud matching and polynomial feature space transformation estimation in the frequency domain using the fast Fourier transform (FFT). A multiscale phase-based feature extraction method is proposed that determines both the location and size of the dynamic sub-clouds being extracted. A simple outlier pruning based on resampling is used to remove false keypoint matches. The proposed phase-based approach to registration can be performed very efficiently without the need for initial estimates or equivalent keypoints from both images. Experimental results show that the proposed method can provide accuracies comparable to the state-of-the-art phase-based image registration methods for the purpose of initial coarse-grained registration while being much faster to compute.