Joint-Saliency-Structure Adaptive Kernel Regression Based Nonrigid Image Registration

Nonrigid image registration has attracted increasing attention in motion tracking, shape comparison, image segmentation and object-based image interpolation for the last decade. However, owing to the outliers introduced by missing correspondences and local large deformations, accurately matching each pair of local structures from those outliers is still a very challenging issue in current nonrigid image registration.To tackle this problem, we defined the nonrigid image registration as a local adaptive kernel regression to locally reconstruct the moving image’s dense deformation vectors from the sparse deformation vectors in the multi-resolution block matching. To gather more sparse deformation vector samples of the same structure, the kernel functions could be compliant with the reference image’s local saliency by adaptively adjusting their orientation and scale in the kernel regression. By these two adaptions, the saliency structures will not change suddenly across the edges and corners so that the local structures in the moving image can preserve their topology in the nonrigid image registration, and thus accurately matching with the local structures in the reference image.To estimate the deformations around the outliers, we use joint saliency map (JSM)[1] that highlights the corresponding saliency structures (called joint saliency structures, JSSs) in the two images to guide the dense deformation reconstruction by emphasizing those JSS’s sparse deformation vectors in the kernel regression. On the other hand, the JSM also marks the bad-aligned regions, i.e. the outliers caused by local large deformation and missing correspondence. Combined the JSM with the scale information of local structures can further promote the adaptive scale selection in JSS adaptive kernel regression.The experimental results demonstrate that by using local JSS adaptive kernel regression, we can accurately match corresponding local structures in the presence of outliers while maintaining an overall smooth deformation field around local structures.