Lukin Krylov Na So Nov
Lukin Krylov Na So Nov
Alexey Lukin, Andrey S. Krylov, Andrey Nasonov Moscow State University, Moscow, Russia [email protected], [email protected], [email protected]
Abstract
Term super-resolution is typically used for a high-resolution image produced from several low-resolution noisy observations. In this paper, we consider the problem of high-quality interpolation of a single noise-free image. Several aspects of the corresponding super-resolution algorithm are investigated: choice of regularization term, dependence of the result on initial approximation, convergence speed, and heuristics to facilitate convergence and improve the visual quality of the resulting image. Keywords: image interpolation, super-resolution, regularization.
image pixels by pixels of the decimated source image. This algorithm provides very good interpolation quality but it is very complex, so it is often executed only in small areas with strong edges while simpler algorithms process the rest of area. The rest of the paper is organized as follows. In section 2, we introduce the super-resolution method for image interpolation and the process of regularization. In section 3, we discuss several variants and modifications of the method and describe their influence on the resulting image with respect to visual artifacts and PSNR quality by giving the results of our experiments. Section 4 concludes the paper by summarizing variants of the superresolution method that provide best image quality.
1. INTRODUCTION
Linear methods for image interpolation are usually constructed to deal with bandlimited signals. The interpolated one-dimensional signal is defined as: F ' ( x ) =
2. SUPER-RESOLUTION METHOD
+ i =
K(x) is the interpolation filter, h is the sampling step. In a twodimensional case, the interpolation is typically performed separately for each axis. The most popular weight functions are box filter (or nearest neighbor), tent function (or bilinear), ideal lowpass filter, Lanczos filter, Gaussian filter, and bicubic interpolation [1]. For every algorithm which is using linear interpolation there are some typical artifacts: blurriness, ringing effect, and jagged edges [6]. Reduction of one of these artifacts increases the others. As usual, non-linear algorithms are used to scale two-dimensional images with a fixed ratio without constructing continuous image. Interpolated pixel values are calculated as a linear combination of nearest sampled values, but the main difference with the linear interpolation is the variability of coefficients which depend on surrounding pixel intensities. The main idea of gradient algorithms is the fact that directed interpolation along edges results in better interpolation than nondirected linear interpolation. The direction and the intensity of an edge in a point are defined by the local gradient information. One of these algorithms is WADI [2], which is based on the modification of bilinear interpolation. It computes the derivate along the normal to every side of a square formed by four sampled pixels and modifies coefficients of bilinear interpolation in a special way: the side with greater derivative results in smaller coefficients for points of this side. Gradient algorithms are fast in the class of non-linear algorithms and produce better results than linear interpolation; it makes edges less jagged and more realistic. NEDI algorithm (New Edge-Directed Interpolation) is a typical non-linear algorithm, which doubles the resolution of images [3]. It uses the concept of self-similarity. The assumption is that coefficients of linear combination used for destination pixel interpolation are the same as coefficients used for interpolation of source
Super-resolution method is typically used to restore a highresolution image from several low-resolution noisy observations [4]. In this paper, we consider the interpolation of a single image. So, we will formulate the problem as
Ax = y ,
(1)
where x is the unknown high-resolution image (represented as a vector of pixel values), y is the known low-resolution image, and A is the downscaling operator typically consisting of decimation D following a low-pass filtering H:
A = DH
(2)
The choice of the low-pass filtering operator depends on a point spread function of the imaging system that produced the lowresolution image. If the imaging system is unknown we will assume that operator H is a simple box filter.
2.1 Regularization
The equation (1) is generally ill-posed and a small change of the input vector y can cause a huge change of the resulting vector x. For the equation (1), the regularized solution is found as:
x = arg min Ax y
p n
+ F ( x ) ,
(3)
where the first term is called as discrepancy, F (x) is the stabilizer and >0 is the coefficient of regularization [5]. The most popular and universal stabilizer is the Tikhonov functional. It is calculated as a grid approximation of the functional:
F ( x ) = x 2 ,
2
(4)
and n=2, p=2. For each >0 the solution x is correct: it is unique, defined for every y and continuously depends on y. We can write the Euler equation for this case:
( AT A + 2 ) x = AT y .
But in this case the algorithm becomes linear because x is the solution of the system of linear equations. So, this method inherits drawbacks of linear interpolation algorithms and we need to find more adaptive stabilizer for image resampling.
We will consider Total Variation (TV) and Bilateral TV (BTV) stabilizers [4], which are working in l1 norm (n=1, p=1):
TV ( x ) = x 1 ,
where x is the gradient operator (its modulus),
s ,t = p
(5)
proximation (NEDI) helps getting better results after fixed number of iterations. Fig. 1 displays improvement of SNR (ISNR) against bilinear interpolation for the super-resolution method with different initial approximations.
ISNR, dB 3
BTV ( x) =
where S and
s x
s ,t = p
t y
s+t
t x S xs S y x , 1
(6)
2,5 2 1,5 1 0,5 Bilinear interp. SR (x0=bilinear) SR (x0=bicubic) SR (x0=NEDI)
s ,t = p
s ,t = p
(7)
0 Cactus
T
Lena
Lighthouse
Text
Average Im age
z=sign x is a vector with per-element applied sign function; D is an up-scaling operation. If D in (2) is the simplest decimation operator that takes every k-th pixel, DT is the up-scaling operator by zero insertion. If H in (2) is a symmetric filtering, then HT is equal to H. x0 is the initial approximation of the high resolution image.
We have noted that the number of iterations between updates of for a fixed divisor depends very weakly on the source image and iteration number. So, after the first fluctuation, we may use the geometric progression instead of the piecewise constant behavior of (fig. 2). So, we have used the following strategy: a certain number of first iterations are processed with a constant coefficient, and after the beginning of fluctuations, the coefficient is exponentially decreased (geometrical progression). Good selection of the initial approximation image makes it possible to fix the number of iterations with some constant initial and the method becomes independent on the source image. Here are some graphs for different ways of changing . The thick line is , the thin line is the discrepancy, and the dotted line is the mean square error between the current solution x and the ground truth result. The vertical scale is logarithmic.
Divisor = 2 100 16,2 16 15,8 15,6 10 15,4 15,2 15 14,8 14,6 1 1 4 7 10 13 16 19 22 25 28 31 34 37 40 14,4
Strong regularization using Tikhonov stabilizer results in strong blur, weak regularization leaves jagged edges. l1-regularization (such as TV or BTV) is different: strong regularization sharpens the edges, trying to make the image piecewiseconstant [6]. The Gibbs phenomenon is effectively suppressed.
Geometric progression with a factor of 0.875 100 16,2 16 15,8 10 15,6 15,4 15,2 1 1 4 7 10 13 16 19 22 25 28 31 34 37 40 15 14,8 14,6 0,1 14,4
4. 5. 6.
The next improvement of the iterative method is modifying D and H operators. Originally, the upsampling HTDT of the sign operator of the discrepancy H T D T sign( DHx y ) is simple: we interpolate the discrepancy with zeros and apply the filter HT. This is a linear method and it does not preserve edge directions, so it results in some edge jaggedness. We propose to use edgedirectional interpolation (such as gradient interpolation) at this stage to reduce effect of jagged edges. The coefficients of gradient interpolation can be calculated only once, using the first approximation image x0. The gradient interpolation is applied at each iteration to enlarge the modified difference. This modification of the original method increases both PSNR and visual quality. Fig. 3 compares PSNR for bilinear vs. gradient upsampling of discrepancy.
Bilinear interp. Bicubic interp. Gradient interp. NEDI SR (x0=bilinear) SR (x0=NEDI, a=grad) SR (fractal subst.)
Figure 4: Comparison of visual quality for bilinear interpolation (left) and the proposed super-resolution algorithm (right).
Lena Lighthouse Text Average Im age
5. REFERENCES
[1] Ken Turkowski Filters for Common Resampling Tasks // Graphics gems, pp. 147-165, Academic Press Professional, Inc., 1990. [2] Shuai Yuan, Masahide Abe, Akira Taguchi, Masayuki Kawamata High Accuracy WADI Image Interpolation with Local Gradient Features // Proceedings of 2005 International Symposium on Intelligent Signal Processing and Communication Systems pp. 85-88, 2005. [3] J.A. Leitao, M. Zhao and G. de Haan Content-Adaptive Video Up-Scaling for High-Definition Displays // Proceedings of IVCP 2003, vol. 5022, January 2003. [4] Sina Farsiu, Dirk Robinson, Michael Elad, Peyman Milanfar Fast and Robust Multi-Frame Super-Resolution // IEEE Trans. On Image Processing, Vol. 13, No. 10, pp. 1327-1344, October 2004. [5] A. Tikhonov, V. Arsenin Solutions of Ill-Posed Problems // Washington DC: WH Winston, 1977. [6] Demo web page with image examples: http://audio.rightmark.org/lukin/graphics/superres.htm
4. CONCLUSION
We have analyzed properties of super-resolution method applied to the interpolation of a single image. It was shown that modifications of the original super-resolution method given by (7) can improve the visual quality and PSNR of resulting images. The example of work of the algorithm is given on a fig. 4, more examples can be found in [6]. The careful choice of initial approximation prevents jagged edges artifact from happening. We suggest using the NEDI algorithm to calculate the initial approximation image. The gradient-directed interpolation used for upsampling of the difference image prevents jagged edges occurring in the process of iterations. The choice of Bilateral TV regularization (6) doesnt blur the edges while reducing the ringing artifact and limiting the noise amplification. The adaptive way of modification of parameter allows reducing the number of required iterations to 16-20 iterations for producing results with best PSNR. The simulation on a 1 GHz P3 computer has shown that it takes approximately 1 minute to produce a 1-megapixel enlarged image from a quarter-megapixel source image (for 16 iterations).
About authors
Andrey S. Krylov is an associated professor at the Moscow State University, Faculty of Computational Mathematics and Cybernetics. Email: [email protected] Alexey Lukin is a member of scientific staff at the Moscow State University, Faculty of Computational Mathematics and Cybernetics, member of IEEE and AES. Email: [email protected] Andrey Nasonov is a student of the Moscow State University, Faculty of Computational Mathematics and Cybernetics. Email: [email protected] Acknowledgements. Alexey Lukin is grateful to Dr. Steven A. Ruzinsky for fruitful discussions on image interpolation algorithms. This work has been partially funded by the Russian Foundation for Basic Research, grant 06-01-00789.