Image Registration Method Based On Logarithmic Polar Coordinate Feature And Nearest Point Iteration

Images can have substantially change their appearance and shape when they are acquired using different modalities,with lighting variations,rotation/optical zoom,having physical changes of scene or at widely different viewpoints.However,even with the state-of-the-art technology,e.g.,the generalized dual-bootstrap iterative closest point method(GDB-ICP),it is still difficult to register those types of images.To handle those challenging images,the present paper proposes a log-polar feature guided iterative closest point algorithm(LPF-ICP)for image registration.Similar to the GDBICP method,the LPF-ICP image registration proceeds in a two-step(or seed and growth)process.First,the LPF-ICP method forms initial bootstrap regions via seed matching.Next,it expands the regions until both of them cover the entire overlap between images via an feature-driven iterative closest point(ICP)process.However,in each step,the techniques adopted by the LPF-ICP method are different from those by the GDB-ICP method.In the initialization step,the GDB-ICP method extracts,as seed,the scaleinvariant bulb points(or SIFT keypoints)from images in scale space.This causes GDBICP method fails in the case of registration at the time of registration large image distortion.Targeting this issue,the present paper proposes a scale-invariant feature detector in log-polar space without loss of any information.By using this kind of LP feature(including corner and bulb)points as seed,the LPF-ICP method derives the initial similar transformation via matching of LP seeds and form the initial bootstrap regions as well.In the growth step,the LPF-ICP method simply uses the singlescale features from the original images instead of the multiscale approach to drive the ICP,thus,gradually expanding the bootstrap regions and refining the transformation estimate.The performance,measured by matching-success rate,alignment error and running time,of the LPF-ICP method is evaluated in registering Rensselaer data set,which includes which includes 22 different image pairs with large variations of appearance,by comparing with the GDB-ICP method.The obtained results show that the matching-success rate of LPF-ICP method achieves 100% match=success rate,while the GDB-ICP has 86%,which,thus,demonstrate that the former method apparently outperforms the later.Next,the average of alignment error for the LPF-ICP and the GDB-ICP both reach to sub pixel and comparable to each other.The present work shows that the LPCD method can reliably extract the LP corner points from images with large appearance changes.Furthermore,by matching of the LP seeds(including LP corners and LP bulbs)the initial bootstrap regions can be formed for image registration.Besides,simply by using the singlescale features(including Harris corners and edges)to drive the ICP process,it is practicable to gradually expand the bootstrap regions and to refine the transformation estimate until matching of the entire image.By combining the above two techniques together,the present paper proposes an LP seed-plus-growth approach to image registration,i.e.,IPF-ICP method.Finally,the experimental evaluation on the Rensselaer data set illustrates that the LPF-ICP method succeeds in all 22 image pairs,while the GDB-ICP complete 19 of them,thus,verifying the effectiveness of the proposed method.