1: \begin{abstract}

2: The filling-in effect of diffusion processes is a powerful tool for

3: various image analysis tasks such as inpainting-based compression 

4: and dense optic flow computation. 

5: For noisy data, an interesting side effect occurs: The interpolated 

6: data have higher confidence, since they average information from many 

7: noisy sources. This observation forms the basis of our denoising 

8: by inpainting (DbI) framework. It averages multiple inpainting results 

9: from different noisy subsets. Our goal is to obtain fundamental insights 

10: into key properties of DbI and its connections to existing methods.

11: Like in inpainting-based image compression, we choose homogeneous 

12: diffusion as a very simple inpainting operator that performs well 

13: for highly optimized data. We propose several strategies to choose the 

14: location of the selected pixels. Moreover, to improve the global 

15: approximation quality further, we also allow to change the 

16: function values of the noisy pixels. In contrast to traditional 

17: denoising methods that adapt the operator to the data,

18: our approach adapts the data to the operator. 

19: Experimentally we show that replacing homogeneous diffusion inpainting

20: by biharmonic inpainting does not improve the reconstruction quality. 

21: This again emphasizes the importance of data adaptivity over operator

22: adaptivity. On the foundational side, we establish deterministic and

23: probabilistic theories with convergence estimates.

24: In the non-adaptive 1-D case, we derive equivalence results between 

25: DbI on shifted regular grids and classical homogeneous diffusion 

26: filtering via an explicit relation between the density and the diffusion 

27: time. 

28: \end{abstract}

29: