24-09-2010, 04:24 PM
IMAGE IMPAINTING BY PATCH PROPAGATION USING PATCH SPARSITY
Presented By:
Tiya Cyriac
S7 AE
Roll No:68
College Of Engineering , Trivandrum
2007-11 batch
OVER VIEW
Introduction .
Algorithm overview.
Inpainting by patch sparsity.
Experiments and comparison.
Conclusion.
IMAGE INPAINTING
Filling of missing region of an image.
Used widely in the field of computer vision & image processing.
Wide applications of digital effects(eg:object removal),image restoration(eg: scratch or text removal in photograph),image coding& transmission(recovery of missing blocks).
DIFFUSION BASED APPROACH
Missing region filled by diffusing image information from known region at pixel level.
Based on theory of PDE .
Isophotes propagated into missing regions.
Smooth effect in larger missing regions.
EXAMPLAR BASED APPROACH
Propagates image information from known region at patch level.
Based on texture synthesis technique.
Texture synthesized by sampling best match patch from known region.
Plausible results for filling large missing regions.
PROPOSED APPROACH
Examplar based image inpainting through patch propagation.
Basic procedures of patch propagation- patch selection and patch inpainting.
Patch selection :- patch on missing region boundary, highest priority.
Patch inpainting :- best match patch from a set of candidate patches for robust inpainting.
PATCH SPARSITY
Patches distributed sparsely over a region.
Two concepts of patch sparsity.
Patch structure sparsity.
Patch sparse representation.
PATCH STRUCTURE SPARSITY
Sparseness of patch’s nonzero similarities with neighbouring patches.
Greater for patch on structure than texture.
PATCH SPARSE REPRESENTATION: Missing patch represented as a linear combination of candidate patches.
ALGORITHM OVERVIEW
INPAINTING USING PATCH SPARSE REPRESENTATION
CONSTRAINTS
INPAINTING EXAMPLES
PATCH SELECTION
Number of non-zero components varies for each patch
SCRATCH AND TEXT REMOVAL
Inpainting algorithm can be used to remove scratches and text on an image.
PSNR values between original and inpainted image, used for quantitative comparison.
IMPLEMENTATION
Small subset of candidate patches decreases computational overhead.
Takes 103 secs to fill missing region with 5310 missing pixels using C++ programming language on intel 2 GHz cpu.
CONCLUSION
IIP algorithm can be used for text removal, object removal, missing block completion.
Two types of patch sparsity proposed.
Patch with larger structure sparsity at structure inpainted with higher priority.
Human labeled structures can be incorporated to recover totally removed structures in future.
REFERENCES
A. Wong and J. Orchard, “A nonlocal-means approach to examplar based inpainting,” presented at the IEEE Int. Conf. Image Processing,
2008.
B. Shen, W. Hu, Y. Zhang, and Y. Zhang, “Image inpainting via sparse
representation,” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal
Processing, 2009, pp. 697–700.
J. C. Yang, J. Wright, T. Huang, and Y. Ma, “Image super-resolution
as sparse representation of raw image patches,” presented at the
IEEE Computer Society Conf. Computer Vision and Pattern Recogition, 2008
M. J. Fadili, J. L. Starck, and F. Murtagh, “Inpainting and zooming
using sparse representations,” The Comput. J., vol. 52, no. 1, pp. 64–79,
2009.
dowload the ppt here:
http://www.filesonicfile/21413165/IMAGE IMPAINTING_2003.ppt