You are on page 1of 5

NOISE ROBUST SINGLE IMAGE SUPER-RESOLUTION USING A MULTISCALE IMAGE

PYRAMID

Jing Hu*, Jiaxin Li, Xi Wu, Jiliu Zhou

Department of Computer Science, Chengdu University of Information Technology, P.R. China, 610225

ABSTRACT already existed in the original LR image can be enhanced.


Therefore, a fundamental problem of SLSR is how to
Single image super-resolution (SR) generates a high- enhance HF details properly. Inappropriate HF enhancement
resolution (HR) image by estimating the mapping function is not only at the risk of producing overshoot artifacts
between image patches of different resolutions. However, around edges in SR result, but also amplifies LR image noise
this kind of SR method cannot be directly applied to noisy as well. However, although several researchers [8, 9] have
images, since noise will be reinforced in the process of pointed out the noise amplification problem of SLSR
super-resolution. To this end, this paper presents a framework, surprisingly, to the best of our knowledge, a
simultaneous super-resolution and denoising method by very limited number of SLSR works have seriously
exploiting the noise decreasing property contained in the addressed this issue. Typically, when dealing with noisy
multiscale image pyramid. Experimental results confirm that images, most SLSR methods involve a two-step process:
our method is able to outperform other state-of-the-art super- super-resolution and denoising. Unfortunately, Singh et al.
resolution methods when super-resolving noisy images [10] have acclaimed that such two-step SLSR framework
across differing noise levels. would reinforce the spatial correlation of the original image
Index Terms—single-image super-resolution, noise- noise by the initial SR process, and the following denoising
robust, multiscale image, self-similarity step cannot remove such noise effectively.
To this end, a noise-robust SLSR framework is studied
1. INTRODUCTION in this paper. Based on observation that the interpolation
process embedded in conventional SLSR accounts for noise
Single image super-resolution (SR) aims to generate a high amplification, we propose a simultaneous interpolation and
resolution (HR) image from a single input low resolution denoising method and combine it to the SLSR framework.
(LR) image. The key problem of SR is to synthesize the Fig. 1(b) is the flowchart of the proposed method.
high-frequency (HF) details which are lost in the LR image. The remainder of this paper is structured as follows. In
Some single image SR methods [1-4], or called learning- Section 2, we first briefly introduce the SLSR framework
based SR methods, learn such details from a set of external and then illustrate the proposed simultaneous interpolation
LR-HR patch pairs, while others[5-8], named as self- and denoising method. Experimental results on several noisy
learning-based SR (SLSR) methods, use the LR input image images are provided in Section 3, together with visual and
itself as the only source of LR-HR example patches. The quantitative comparisons with other methods. Finally, the
common idea of both types of SR approaches is more or less conclusions are described in Section 4.
the same. That is, to infer an HR image or HR patch by
learning the relationship between LR-HR exemplar patch 2. METHOD
pairs. However, utilizing the external training images,
learning-based SR methods are very likely to introduce We develop our combined SLSR method based on our
irrelevant HF details if the LR input image is incompatible previous work on SLSR [11]. We firstly introduce the SLSR
with the training datasets [1, 5]. On the contrary, based on framework and explain the problem of directly applying
the observation that local image structures tend to reappear such framework to noisy LR image. Then we summarize our
within and across different image scales [7], SLSR observation and propose our combined SR framework to
approaches regularize the HF detail estimation problem on address this problem.
these self-examples, and therefore circumvent the false- For convenience, the variables used in the proposed
detail issue. method are defined as follows: as shown in Fig. 1(a),
Although SLSR avoids false-detail problem, its matrices I (I0) and H are used to represent the input LR
fundamental drawback is that only the HF details that have image and the output HR image, I0,s is the smoothed version

978-1-4799-7061-2/18/$31.00 ©2018 IEEE 2526 ICIP 2018


Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 17,2023 at 06:35:58 UTC from IEEE Xplore. Restrictions apply.
of I0 and I1,s is an interpolated version of I0. The subscript t q  f qs   f ps  qs  ps 
in I   denotes t-th iteration of the algorithm. The bolded
t
1 (1)
 f  p s   f   p s   q s  p s   f   p s   q s  p s   q s  p s  
lowercase p and q terms denote the column vectors of two 2
s×s image patches which are extracted from I 0  and I1  ,
t t
where denotes the element-wise product, f    and
respectively; ps and qs are the column vectors of two s×s f    denote the first- and second derivatives of the
image patches taken from I 0,t s and I1, t s , respectively. Without regression function. Moreover, by assuming the input LR
specific notification, we assume that p, ps, q, and qs are image is noise-free and the mapping function follows some
related. That is, ps is the most similar among all patches to smoothness property, we come to p = f(ps). Therefore, Eq.
qs in the image I 0,t s ; patches p (q) and ps (qs) have the same (1) is reformulated as:
1
coordinates for the center pixel. {ps, p} constitutes the LR- q  p  f   p s   q s  p s   f   p s   q s  p s   q s  p s  (2)
HR training patch pairs and {qs, q} constitutes the LR-HR 2
testing patch pairs. Intuitively, to obtain HR patch q , the derivatives of
I1 
1

1
I1,s mapping function should be firstly obtained. To achieve this,
( H ) we use the algorithm proposed in our previous work that
exploits the multiscale non-local similarity property to select
f
suitable exemplar LR-HR training image patches and
achieves an accurate second-order derivative estimation for
the mapping function. More specifically, patches (colored in
 
I1,s
0 green in Fig. 1(a)) similar to ps and their corresponding HR
I1 
0

1
( I0 ) version (colored in pink in Fig. 1(a)) are used to estimate the
q
f qs
mapping function f that associates ps to p. For more
technical details, the interested reader may refer to [11].

Unified interpolation and denoising

I 0,s
0

Input I p ps
( I 00 )
f

I 1,s
0
 0
I 1
(a) (b) (c) (d)

Fig. 2. Controlled experiments at different settings of p, f’(ps), f’’(ps), and


I 2  0
0
I 2,s qs.. The input LR image was contaminated with Gaussian noise (σ = 10). (a)
LR input image (enlarged using nearest-neighbor interpolation for better
comparison). (b)-(d) are SR results using the second-order-based SLSR
Mapping function f estimation
framework, where (b) p, f’(ps), f’’(ps), and qs. are both noisy, (c) p, f’(ps),
f’’(ps) is noise-free and qs is noisy, (d) p, f’(ps), f’’(ps) are noisy and qs is
(a)
I 0
iter 
noise-free.

f
I (0iter 1) Mapping function Unified interpolation
Image
Final
Output
However, the second-order-based SLSR only works
estimation and denoising
resize H well on noise-free images. Several researchers [8,11] have
pointed out that directly applying SLSR to noisy LR image
(b) would cause noise amplification in the reconstructed HR
Fig. 1. An SLSR flowchart. (a) SR reconstruction inside each iteration and image, as shown in Fig. 2(b). Analytically, if the input image
(b) iteration process. Note that patches p (q) and ps (qs) have the same is noisy, these four terms in Eq. (2) would also be corrupted:
ordinate. p, f’(ps), f’’(ps), and qs. To investigate the term that is the
most essential factor for noise amplification in SLSR, two
2.1 Second-order derivative estimation based SLSR controlled experiments were carried out. In the first
framework empirical experiment, we assume that the ground truth of p,
f’(ps), f’’(ps) are known, which is estimated from the original
SLSR assumes that mapping function f associates each LR- noise-free LR image, whereas qs is calculated from the noisy
HR patch pair {qs, q}as q = f(qs). By applying an Nth order LR input. In the second empirical experiment, we assume
Taylor series, the unspecified f can be locally expanded in ps, that the interpolated image is calculated from the original
if ps and qs are similar in terms of the geometric layout: noise-free image, and other three terms are estimated from
the noisy input. Figs. 2(c) and 2(d) represent the
corresponding reconstruction results under these

2527
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 17,2023 at 06:35:58 UTC from IEEE Xplore. Restrictions apply.
assumptions. Obviously, we can observe from Fig. 2 that represent neighboring pixels used in conventional interpolation and the
pink line show the “best” neighboring pixels used in the proposed method
among these four terms, the noise in qs is more influential
than other three in noise amplification. Therefore, how to In our method, three types of directional multiscale
remove the noise in qs, or in other words, the interpolated pyramids are exploited. More specifically, they are
version I1,s in SLSR, is fundamental for a noise-robust generated by blurring and downsampling I0 in horizontal,
SLSR algorithm. vertical, and diagonal directions. Corresponding scaling
factors of d n , n  1, 2, , 20 are used. For pyramids in the
2.2 Simultaneous interpolation and image denoising
horizontal and vertical directions, variance of the 5×5
Generally, there are two different ways in getting a clean I1,s Gaussian blur kernel in each scale was set to  n2  d n n ,
from the noisy LR input: performing denoising and where σ denotes the standard deviation of noise in an input
interpolation (bicubic in our method) separately or LR image. For the diagonal pyramid, the variance of the
simultaneously. Our pilot experiments on BSD500 database 5×5 Gaussian blur kernel at each scale was set
reveal that the former way would weaken fine image details to  n2  d n B2 n . Then, we empirically look for “best”
(see Fig. 3 for details). Therefore, we propose a joint candidate of each P0 from these directional pyramids. As
interpolation and denoising procedure. shown in Fig. 4, the candidates are located at the same
relative image coordinates to P0 in three directional pyramid.
For instance, if (j, k) are the coordinates of P0’s centered
pixel, its candidate in the nth down-scale is located near the
coordinates (|j×dn|, |k×dn|), n = 1, 2, …, 20. In this way, we
are able to form a total of 60 (3×20) candidate patches for
(a)
each noisy patch.
(b) (c) (d) (e) Intuitively, the best candidate for P0 should account for
Fig. 3. Example of our pilot study on the order of denoising and structure preservation of the noisy input and noise-level
interpolation. (a) LR image with σ = 10, (b)-(d) are results of our reduction. In this paper, these two factors are associated via
proposed framework using (b) simultaneous interpolation and denoising,
(c) interpolation+denoising, (d) denoising+interpolation. (e) Original HR
the following objective function:
image. 1 1
min dist  P0 , P- n   min 2 P0  P- n 2   2 d n 1
2
(3)
Observed that the noise level becomes lower along P- n P n N 2
with the decreased image scale in the multiscale image where P0 denotes the noisy input patch in I0 and P-n is the
pyramid [12], this multiscale noise reduction prior is in-scale candidate patch in the n-th down-scale image (both
employed in our interpolation method. Considering that in a P0 and P-n are normalized to [0, 1]), N represents the pixel
typical image interpolation scenario, a target pixel is number within P0. The first term in Eq. (3) represents the
estimated by a weighted averaging of its neighboring pixels, data fidelity and the second term approximates the noise
in which the interpolation weight is determined by the correlation between an original noisy patch and the
spatial distance between the target pixel and its neighboring multiscale candidate.
pixel. Thus, in order to achieve a noise-free interpolation, After selecting the optimal candidate, it is then used as
we propose to specially select the neighboring pixels with the neighboring pixels for estimating the target pixel y in the
lower noise level from a multiscale directional pyramid. interpolated I1,s, as shown in Fig. 4. In a similar fashion, all
I1

pixels in image I1,s are estimated in a raster-scan order and


I0
P0
the noise-free interpolated image I1,s is consequently
y
obtained. By substituting the joint interpolation and
denoising procedure with simple interpolation of the SLSR
I-1 P-1
framework in 2.1, we can obtain the noise-robust SLSR
method.
I-2
P-2

I-3
P-3 3. EXPERIMENT
P-4
I-4
In order to evaluate the robustness of the proposed method,
… P-14 several experiments were conducted on noisy images with
I-14


different noise levels. 26 typical images are used as the
P-20
I-20
ground truth for synthesis experiments. Several state-of-the-
Fig. 4. The diagram of the proposed joint denoising and interpolation. art SR approaches including the Kim’s method [4], deep-
Patches Pi i = 0,-1,…,-20 have the same relative coordinate. Note that only learning-based SR (DLB-SR) [13], the A+ method [14],
the diagonal pyramid is shown for display convenience. The green line
MTR method [15], and the NCSR method [9] were selected

2528
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 17,2023 at 06:35:58 UTC from IEEE Xplore. Restrictions apply.
TABLE I Comparisons of average PSNR values [dB] (the first row) and SSIM values (the second row)
for images with different noise levels

Kim+ A++ DLB-SR+


σ Kim A+ DLB-SR MTR NCNR Ours
denoise denoise denoise
25.86 27.59 25.95 27.49 25.84 27.61 24.64 26.55 27.72
10
0.66 0.77 0.66 0.77 0.65 0.79 0.72 0.75 0.82
23.65 26.45 23.77 26.49 23.57 26.50 24.27 25.23 26.65
15
0.54 0.74 0.55 0.74 0.53 0.73 0.68 0.65 0.77

as comparison baselines. The latter two are noise-robust from σ = 10 to σ = 15, the gap in PSNR values between
methods. Note that Kim’s method, DLB-SR and the A+ NCSR and the denoised version of A+ is also increased.
method cannot handle noisy images directly. An advanced This is partly due to images which are seriously
image denoising method PGPD [16] is leveraged to further contaminated, the noise-robust method NCSR applies as an
process their SR results. LR images used in our experiments averaging technique and cannot effectively remove noise.
are generated by down-sampling (bicubic interpolation) the On the contrary, the proposed method consistently
original images, followed by corruption of white Gaussian outperforms other SR methods as well as the two-step
noise with zero mean and standard deviation σ. framework across all test images. This serves to validate the
In the proposed method, the patch size was set to 5×5, effectives of the proposed method in processing of noisy
with 4 pixels overlapping between adjacent patches. A images.
standard deviation σ equals to the noise level deviation was For quantitative comparison, the 2× SR results for noisy
assigned to the Gaussian filters used to generate the image Flower with a noise standard deviation of 15 are
multiscale directional pyramids. In this paper, the noise shown in Figs. 5. For better visual comparison, we also
level was estimated by the method advocated in [17]. In provide local magnification in the blue rectangle in the
addition, the gradual upscaling factor was set to 1.5, and the lower-right corner in each example image. It is observed
down-scaling factor d was set to 0.9. that during the SR process, additive noise in the LR image
Table I summarizes the average PSNR and SSIM [18] is also introduced to the output image. Although a post
values for the abovementioned SR algorithms on LR images denoising step is able to remove most noise components, the
with varying noise levels (σ = 10, and 15). We can see that magnitude of remnant noise cannot be neglected. As can be
the PSNR and SSIM scores in Kim’s method, DLB-SR and seen from these images, the proposed method not only
the A+ method are lower than the noise-robust methods. provides a better noise removal than any of the compared
This is because these three methods lack a noise attenuation algorithms, but also produces fewer jagged or blurry
technique and noise amplitude is magnified during artifacts.
superresolving. Moreover, we observe that after the
denoising step, the PNSR and SSIM values in these three
methods are increased above that of the MTR and NCSR 4. CONCLUSION
methods. This confirms that the noise issue in SR cannot be
neglected. More specifically, as the noise level increases
In this paper, we presented a new SLSR method for
processing noisy images. The proposed method first
explored the cause of noise magnification in the SLSR
framework and then developed a simultaneous
(a) (b) (c) (g)
interpolation and denoising scheme by using the noise
reduction property of multiscale image pyramid.
Substituting conventional interpolation method with this
scheme in the SLSR framework, a noise-robust SLSR is
developed. Extensive experimental results obtained from
different noisy images have validated the effectiveness of
(h) (i) (j) (k) the proposed SR method for detail preservation and noise
removal. The proposed technique produced an improved
Fig. 5. A comparison of the SR results for the noisy image FLower with super-resolution reconstruction both quantitatively and
σ = 15. (a) A+. (b) Denoised A+ result. (c) NCSR. (g) DLB-SR. (h)
perceptually in comparison with other advanced baselines.
Denoised DLB-SR. (i) The proposed method. (j) The LR image. (k) The
ground truth. Zoom for better comparison

2529
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 17,2023 at 06:35:58 UTC from IEEE Xplore. Restrictions apply.
REFERENCES [10] A. Singh, F. Porikli, and N. Ahuja, “Super-resolving noisy
images,” in Proc. IEEE Comput. Soc. Conf. on Computer
[1] W. T. Freeman, T. R. Jones, and E. C. Pasztor, “Example- Vision and Pattern Recognition, pp. 2846-2853, (2014).
based super-resolution,” IEEE Comput. Graph. Appl. 22(2), [11] J. Hu and Y. Luo, “Single-image superresolution based on
56-65 (2002). local regression and nonlocal self-similarity,” Journal of
[2] J. Yu et al., “A unified learning framework for single image Electronic Imaging, 23(3), 033014-033014-14, (2014).
super-resolution,” IEEE Trans. Neural Netw. Learn. Syst. [12] M. Zontak and M. Irani, “Internal statistics of a single
to appear (2013). natural image,” in Proc. IEEE Comput. Soc. Conf. on
[3] X. Gao et al., “Joint learning for single-image super- Computer Vision and Pattern Recognition, pp. 977-984
resolution via a coupled constraint,” IEEE Trans. Image (2011).
Process. 21(2), 469-480 (2012). [13] Z. Wang, et al. "Deep networks for image super-resolution
[4] K. Kim and Y. Kwon, “Single-image super-resolution with sparse prior." in ICCV, pp. 370-378, (2015).
using sparse regression and natural image prior,” IEEE [14] R. Timofte, V. De Smet, and L. Van Gool, “A+: Adjusted
Trans. Pattern Anal. Mach. Intell. 32(6), 1127-1133 (2010). Anchored neighborhood regression for fast super-
[5] G. Freedman and R. Fattal, “Image and video upscaling resolution,” In Asian Conference on Computer Vision
from local self-examples,” ACM Trans. Graph. 30(2) (ACCV 2014), pp.1-5, (2014).
(2011). [15] E. Lopez-Rubio, “Super-resolution from a single noisy
[6] M. Zontak and M. Irani, “Internal statistics of a single image by the median filter transform,” SIAM Journal on
natural image,” in Proc. IEEE Comput. Soc. Conf. on Imaging Sciences, 9(1), 82-115 (2016).
Computer Vision and Pattern Recognition, pp. 977-984 [16] J. Xu, et al., “Patch group based nonlocal self-similarity
(2011). prior learning for image denoising,” in Proc. of IEEE Int.
[7] K. Zhang, X. Gao, D. Tao, and X. Li, “Single image super- Con. on Computer Vision, pp. 244-252, (2015).
resolution with multiscale similarity learning,” IEEE Trans. [17] X. Liu, M. Tanaka and M. Okutomi, “ Noise level
Image Process. 24(10), 1648-1659 (2013). estimation using weak textured patches of a single noisy
[8] J. Yang, Z. Lin, and S. Cohen, “Fast Image super- image,” in Proc. of IEEE Int. Con. on Image Processing, pp.
resolution based on in-place example regression,” in Proc. 665-668, (2012).
IEEE Comput. Soc. Conf. on Computer Vision and Pattern [18] Z. Wang et al., “Image quality assessment: from error
Recognition, pp. 1059-1066, (2013). visibility to structural similarity,” IEEE Trans. Image
[9] W. Dong, et al, “ Nonlocally centralized sparse Process. 13(4), 600-612 (2004).
representation for image restoration,” IEEE Transactions
on Image Processing, 22(4), 1620-1630, (2013).

2530
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 17,2023 at 06:35:58 UTC from IEEE Xplore. Restrictions apply.

You might also like