“Filling-in stochastic texture” by Sakaino

We present a dynamic patch (DP) method for filling-in a hole of a stochastic/dynamic texture image using a Graph Cut (GC) [2], the advection equation (AE), and optical flow (OF) database. The GC method is iteratively applied with spatio-temporally changing the patch texture over time. This synthesizes naturally changing texture inside and boundary of the filling hole. Previously, an exemplar-based [3,4] and diffusion equation [1] based method have been reported. Issue: Most of them failed to non-periodically fill in dynamic texture/stochastic pattern [6] such as ocean and cloud when a gap becomes wide. Only copying of neighboring similiarity from a source image to a hole limits this because of no dynamic change in an inner patch, i.e., essentially static. An incomplete seamless boundary is synthesized by the previous methods [4]. A trial to transform an inner patch by such as rotation did not solve this issue. One possibility to solve this is to change the topology of a patch like a flow-guidance [3]. We extended this idea.

1. M. Bertalmio et al. 2001. Navier-stokes, fluid dynamics, and image and video inpainting. IEEE CVPR, 1, pp. 355–362.
2. V. Kwatra et al. 2003. Graphcut Textures: Image and video synthesis using graph cuts. ACM Proc. SIGGRAPH.
3. V. Kwatra et al. 2005. Texture optimization for example-based synthesis. ACM Proc. SIGGRAPH.
4. A. Criminisi et al. 2004. Region filling and object removal by exemplar-based image inpainting. IEEE Trans. Image Process
5. J. Stam. 1999. Stable fluids. ACM Proc. SIGGRAPH.
6. Yung-Yu et al. 2005. Animating pictures with stochastic motion textures. ACM Proc. SIGGRAPH, 24, 3.