Gauss-Newton image registration with CUDA

Image registration is the process of matching different images whether 2D or 3D of certain similar or common properties for different purposes. This work addresses this field using a Gauss-Newton optimization approach. The problem is basically formulated as minimizing a cost function that is then solved by a backtracking line search. Since this is considered as a demanding problem especially for larger data, this paper presents the solution using the CUDA GPU architecture provided by Nvidia [1] in order to achieve better performance and reduce timing through parallelism.