Random walker image registration with cost aggregation
Ying Tang, Ghassan Hamarneh · 2014
Random walker image registration (RWIR) [1] has recently been shown to be a promising method for deformable registration. The computational complexity of RWIR, however, depends on two factors: 1) the number of unknown variables, and 2) the number of discrete values each variable can take. Having large values for both factors increases flexibility in the spatial transformation model, which permits fine features in the deformation fields, thereby allowing for increased accuracy in the solution. However, this comes at a cost of increased memory requirements and computation. In this paper, we propose cost aggregation for RWIR to reduce the number of unknowns, such that we group nodes with similar data costs and obtain a non-uniform discretization of the image domain, through which we perform registration. As our experimental results show, we are able to achieve comparable registration accuracy than that achieved by a uniform discretization [1] while lowering complexity.