MARKOV RANDOM FIELD BASED IMAGE INPAINTING WITH CONTEXT-AWARE LABEL SELECTION
These are the results accompanying the paper "Markov Random Field based image inpainting with context-aware label selection" submitted to the International Conference on Image Processing (ICIP) 2012.
We show separately results of:
Method 1 - the proposed efficient optimization method (explained in Sec. 3.2 of the paper) without context-aware label selection
Method 2 - complete proposed method, with label selection and efficient optimization.
We compared the proposed method with following state-of-the-art methods:
[1] Komodakis, N., Tziritas, G.: Image completion using efficient belief propagation via priority scheduling and dynamic pruning. IEEE Trans. Image Proc., vol. 16, no. 11, 2649-2661 (2007)
[2] Le Meur, O., Gautier, J., Guillemot, C.: Examplar-based inpainting based on local geometry. In ICIP 2011, pp. 3462-3465 (2011)
Comparison of global inpainting methods
Input image
![](bungeeMasked.png) |
|
Result of [1]
![](bungee_origMRF.png) |
|
Result of proposed Method 1
![](bungee_simple.png) |
|
Result of proposed Method 2
![](bungee_proposed.png) |
|
Input image
![](vegasMasked.png) |
|
Result of [1]
![](vegas_origMRF.png) |
|
Result of proposed Method 1
![](vegas_simple.png) |
|
Result of proposed Method 2
![](vegas_proposed.png) |
|
Comparison with state-of-the-art methods
Input image
![](snowbaseballMasked.png) |
|
Result of [2]
![](snowbaseball_lemeur.png) |
|
Result of [1]
![](snowbaseball_origMRF.png) |
|
Result of proposed Method 1
![](snowbaseball_simple.png) |
|
Result of proposed Method 2
![](snowbaseball_proposed.png) |
|
Input image
![](elephantMasked.png) |
|
Result of [2]
![](elephant_lemeur.png) |
|
Result of [1]
![](elephant_origMRF.png) |
|
Result of proposed Method 1
![](elephant_simple.png) |
|
Result of proposed Method 2
![](elephant_proposed.png) |
|
Input image
![](sydneyMasked.png) |
|
Result of [2]
![](sydney_lemeur.png) |
|
Result of [1]
![](sydney_origMRF.png) |
|
Result of proposed Method 1
![](sydney_simple.png) |
|
Result of proposed Method 2
![](sydney_proposed.png) |
|
Input image
![](skiresortMasked.png) |
|
Result of [2]
![](skiresort_lemeur.png) |
|
Result of [1]
![](skiresort_origMRF.png) |
|
Result of proposed Method 1
![](skiresort_simple.png) |
|
Result of proposed Method 2
![](skiresort_proposed.png) |
|
Input image
![](soldierMasked.png) |
|
Result of [2]
![](soldier_lemeur.png) |
|
Result of [1]
![](soldier_origMRF.png) |
|
Result of proposed Method 1
![](soldier_simple.png) |
|
Result of proposed Method 2
![](soldier_proposed.png) |
|
Comparison with state-of-the-art method from [2] on images from [3]
[3] Kawai, N., Sato, T., Yokoya, N.: Image inpainting considering brightness change and spatial locality of textures and its evaluation. In PSIVT 2009, pp. 271–282 (2009).
Input image
![](a018Masked.png) |
|
Result of [2]
![](a018_lemeur.png) |
|
Result of proposed Method 2
![](a018_proposed.png) |
|
Input image
![](a036Masked.png) |
|
Result of [2]
![](a036_lemeur.png) |
|
Result of proposed Method 2
![](a036_proposed.png) |
|
Input image
![](a064Masked.png) |
|
Result of [2]
![](a064_lemeur.png) |
|
Result of proposed Method 2
![](a064_proposed.png) |
|
Input image
![](a079Masked.png) |
|
Result of [2]
![](a079_lemeur.png) |
|
Result of proposed Method 2
![](a079_proposed.png) |
|
Dependence of the result of the proposed method (Method 2) on the patch size while keeping the block division fixed (5x7 blocks)
Note that keeping the block division fixed means that the search space is the same for all patch sizes.
Input image
![](elephantMasked.png) |
|
Patch size 7x7
![](elephant_gap3_nBlocks35.png) |
|
Patch size 9x9
![](elephant_gap4_nBlocks35.png) |
|
Patch size 11x11
![](elephant_gap5_nBlocks35.png) |
|
Patch size 13x13
![](elephant_gap6_nBlocks35.png) |
|
Patch size 15x15
![](elephant_gap7_nBlocks35.png) |
|
Patch size 17x17
![](elephant_gap8_nBlocks35.png) |
|
Dependence of the result of the proposed method (Method 2) on the block division while keeping the patch size fixed
Patch size is 11x11 in the first and 15x15 in the second image. Note that the number of blocks along the vertical dimension (M) is the same, but the total number of blocks in two images is different because images are of different sizes and blocks are constrained to be square.
Input image
![](sydneyMasked.png) |
|
4x5 blocks
![](sydney_gap5_nBlocks20.png) |
|
5x7 blocks
![](sydney_gap5_nBlocks35.png) |
|
6x8 blocks
![](sydney_gap5_nBlocks48.png) |
|
Input image
![](vegasMasked.png) |
|
3x4 blocks
![](vegas_gap7_nBlocks12.png) |
|
4x6 blocks
![](vegas_gap7_nBlocks24.png) |
|
5x7 blocks
![](vegas_gap7_nBlocks35.png) |
|
6x9 blocks
![](vegas_gap7_nBlocks54.png) |
|