optimized neighbor embeddings for single-image super-resolution

23
Optimized neighbor embeddings for single- image super-resolution MEHMET TÜRKAN, DOMINIQUE THOREAU, PHILIPPE GUILLOTEL, CHARLY POYAC

Upload: hayes

Post on 23-Feb-2016

61 views

Category:

Documents


0 download

DESCRIPTION

Optimized neighbor embeddings for single-image super-resolution. Mehmet Türkan , Dominique Thoreau, Philippe Guillotel , Charly Poyac. Summary. Basic concepts Principle Quality Improvements Complexity improvements Results Conclusion & remaining work. Basic concepts. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Optimized neighbor embeddings for single-image super-resolution

Optimized neighbor embeddings for single-image super-resolutionMEHMET TÜRKAN, DOMINIQUE THOREAU, PHILIPPE GUILLOTEL, CHARLY POYAC

Page 2: Optimized neighbor embeddings for single-image super-resolution

2

Summary• Basic concepts• Principle• Quality Improvements• Complexity improvements• Results• Conclusion & remaining work

Page 3: Optimized neighbor embeddings for single-image super-resolution

3

Basic concepts

To resize an image, standard algorithms are interpolation based (linear, cubic…), they are fast but they cannot recovery high frequencies and the result is often blur.

Original low res image

Nearest neighbor interpolation

Bicubic interpolation

Page 4: Optimized neighbor embeddings for single-image super-resolution

4

Basic conceptsSuper-resolution is class of techniques that enhance the resolution of an imaging system.

Methods for SR can be broadly classified into two families of methods: The classical multi-image super-resolution[3], and the Example-Based super-resolution[4][5][7][8].

Page 5: Optimized neighbor embeddings for single-image super-resolution

Objectives of the internship

5

Starting from a reference algorithm designed by Mehmet Turkan, Dominique Thoreau, and Phillipe Guillotel, my job was:

• Switch original software developed from Matlab to C/C++.• Speed up the processing time and enhance the quality as much as possible.

Final objective is to upsample FullHD sequences to UHD in a suitable time.

Page 6: Optimized neighbor embeddings for single-image super-resolution

6

Summary• Basic concepts• Principle• Quality Improvements• Complexity improvements• Results• Conclusion & remaining work

Page 7: Optimized neighbor embeddings for single-image super-resolution

7

Principle

Small patches are very likely to repeat themselves within and across different scales of an image.

We can reconstruct a coherent high resolution image with patches correspondence between low res and downsampled images.

Page 8: Optimized neighbor embeddings for single-image super-resolution

8

Principle : minimization

𝐺𝑘=(⟨𝐿𝑅1 ,𝐿𝑅1 ⟩ ⟨𝐿𝑅1 ,𝐿𝑅2 ⟩ … ⟨𝐿𝑅1 ,𝐿𝑅𝑘 ⟩⟨𝐿𝑅2 ,𝐿𝑅1 ⟩ ⟨𝐿𝑅2 ,𝐿𝑅2 ⟩ ⋯ ⟨𝐿𝑅2 ,𝐿𝑅𝑘 ⟩

⋮ ⋮ ⋱ ⋮⟨𝐿𝑅𝑘 ,𝐿𝑅1 ⟩ ⟨𝐿𝑅𝑘 ,𝐿𝑅2 ⟩ ⋯ ⟨𝐿𝑅𝑘 ,𝐿𝑅𝑘 ⟩)h𝑊𝑒𝑖𝑔 𝑡𝑠𝑘=

𝐺𝑘− 1  1

1𝑇𝐺𝑘−1   1

LLE (locally linear embedding) is computed with low res patches to approximate the input patch and weights obtained are used to merge high res patches.

The best K patches combination is obtained by a downsampling of the merged patch and a SSD comparison with the input LR patch.

With the knn-search, we have K couples of low res and high res patches.

Page 9: Optimized neighbor embeddings for single-image super-resolution

9

Page 10: Optimized neighbor embeddings for single-image super-resolution

10

Summary• Basic concepts• Principle• Quality Improvements• Complexity improvements• Results• Conclusion & remaining work

Page 11: Optimized neighbor embeddings for single-image super-resolution

11

Quality Improvements: patch orientation

To increase the result of the knn-search, it’s possible to seek the input patch plus his transformations (transpose, symmetric X, symmetric Y and combinations).

Result is better but the knn-search is about 8 times slower.

Page 12: Optimized neighbor embeddings for single-image super-resolution

12

Quality Improvements: textural features

standard texturalWe can improve the opportunity of find similar patches in the knn-search by using a textural version of all the patches.

In the same way, the LLE is computed with textural patches, and the mean value of the input patch is added to the result because the upsampling of the mean is itself.

Page 13: Optimized neighbor embeddings for single-image super-resolution

13

Quality Improvements: attempts• Using a bicubic neighborhood instead of padding for the patch downsampling.

• Compute a first pass with a big value of K, and a second pass with k varying and using the neighborhood of the first pass.

• Downsampling result and compare it with the input image to detect error areas in order to recompute with better parameters.

• Continue to minimize the patch combination until it’s over a maximum error.

• Nonlinear minimization instead of LLE.

Page 14: Optimized neighbor embeddings for single-image super-resolution

14

Summary• Basic concepts• Principle• Quality Improvements• Complexity improvements• Results• Conclusion & remaining work

Page 15: Optimized neighbor embeddings for single-image super-resolution

15

Complexity improvements: windows search

Instead of the full phases surface, it’s possible to significantly speedup the knn-search using a window.

Results are close to a full search because corresponding patches are often beside the input patch position.

Page 16: Optimized neighbor embeddings for single-image super-resolution

16

Complexity improvements: single channel SR

Because the Human visual system is principally sensible in luminance, it’s possible to convert an image from RGB to YUV colorspace and work only in Y channel, U and V channels are interpolated with Lanczos algorithm.

Visually, results are very close to a multi channel processing but the compute time is faster.

HR image

LR image channels

Page 17: Optimized neighbor embeddings for single-image super-resolution

17

Complexity improvements: minimization

Histogram of best value of K

Standard minimization : K takes all value until

Optimized minimization, only small values of k are computed + max value

→ Result is very close to a full minimization with all values of K.

Number of patches

K

Page 18: Optimized neighbor embeddings for single-image super-resolution

18

Summary• Basic concepts• Principle• Quality Improvements• Speed improvements• Results• Conclusion & remaining work

Page 19: Optimized neighbor embeddings for single-image super-resolution

19

Results : Upsampling x4 for Oldman

LANCZOS OUR[8]ORIGINAL NEAREST NEIGHBOR

Page 20: Optimized neighbor embeddings for single-image super-resolution

20

LANCZOS OUR ORIGINAL

Page 21: Optimized neighbor embeddings for single-image super-resolution

21

Summary• Basic concepts• Principle• Quality Improvements• Complexity improvements• Results• Conclusion & remaining work

Page 22: Optimized neighbor embeddings for single-image super-resolution

22

Conclusion & Remaining work

Remaining work:• CUDA implementation:

Knn-search, matrix inversion and patch combination are very parallelizable.

• Discontinuity problems with sequences: Lot of artifacts appear, especially on edges.

Advantages:• Results are excellent compared to standards interpolation methods (bicubic, Lanczos…), average

is 1.5dB better, many details appear.• It’s visually better than all other super resolution algorithms.

Drawbacks:• Very slow to compute, about 1 hour to upsample a FullHD image into UHD, but it’s 20 times

faster than the reference Matlab software.

Page 23: Optimized neighbor embeddings for single-image super-resolution

23

References1. H. Chang, D.-Y. Yeung, and Y.Xiong, “Super-resolution through neighbor embedding,” in IEEE Comp.

Soc. Conf. Comp. Vis. Pattern Recogn., 2004, pp. 275-282.2. W. T. Freeman, E. C. Pasztor, and O. T. Carmichael, “Learning low-level vision,” Int. J. Comp. Vis., vol.

40, no. 1, pp. 83-97, 2004.3. J. Yang, J. Wright, T. Huang, and Y. Ma, “Image super-resolution via sparse representation,” IEEE

Trans. Image Process., vol. 19, pp. 2861-2873, 2010.4. D. Glasner, S. Bagon, and M. Irani, “Super-resolution from a single image,” in IEEE Int. Conf. Comp.

Vis., 2009, pp. 349-356.5. G. Freedman and R. Fattal, “Image and video upscaling from local self-examples,” ACM Trans.

Graph., vol. 28, no. 3, pp.1-10, 2010.6. S. Roweis and L. Saul, “Nonlinear dimensionality reduction by locally linear embedding,” Science,

vol. 290, pp. 2323-2326, 2000.7. M. Turkan, D. Thoreau, and P. Guillotel, “Optimized neighbor embeddings for single-image super-

resolution,” To be presented in ICIP 2013.8. M. Turkan, D. Thoreau, and P. Guillotel, “Self-content super-resolution for ultra-HD up-sampling,” in

Proc. European Conf. Visual Media Prod., 2012, pp. 49–58.