| Summary: | Image warping is a popular tool for modeling deformation and motion in digital images. Inversion of such models requires two related operators: adjoint and differentiated image warping. Many applications rely on these operators, often through ad hoc and approximate implementations, which leads to a suboptimal quality and convergence speed, and hinders development of and comparison across different applications.In this work, we present an open-source image warping toolbox called ImWIP (Image Warping for Inverse Problems) that overcomes these issues. It implements differentiable image warping operators, together with their exact adjoints and derivatives (up to floating point errors). ImWIP is demonstrated on examples from X-ray computed tomography and magnetic resonance imaging, and is shown to improve both reconstruction quality and convergence speed compared to state-of-the-art warping methods.
|
|---|