International Journal of Engineering and Technical Research (IJETR)

ISSN: 2321-0869, Volume-1, Issue-8, October 2013

Wavelet Transform Techniques for Image

Resolution Enhancement
Prasanth C.R, Sreeja K.S

Abstract I m a g e s are being used i n many fields

of research. One of the major issues of images is their
resolution. In this paper we are studying different image
resolution enhancement techniques that use Wavelet
Transform (WT).
Basis functions of the WT are small waves located in
different times. They are obtained using scaling and
translation of a scaling function and wavelet function
Therefore, the WT is localized in both time and frequency. In
this paper we are comparing different image resolution
enhancement techniques those using Wavelet Transform.

Index Terms Image Interpolation, Peak signal-to-noise Fig 1. Block diagram of DWT Filter Banks of level 1
ratio (PSNR), Wavelet Zero Padding (WZP), Cycle Spanning [8]
(CS), Dual-Tree Complex Wavelet Transform (DT-CWT),
Discrete Wavelet Transform (DWT), Stationary Wavelet
Transform (SWT).
II. REGULARITY-PRESERVING IMAGE
INTERPOLATION
I. INTRODUCTION Traditional interpolation methods work in the time domain.
As stated in [2], the regularity-preserving interpolation
Resolution has been frequently referred as an important technique synthesizes a new wavelet subband based on the
property of an image. Images are being processed in order to known wavelet transform coefficients decay. The lowpass
obtain super enhanced resolution. One of the commonly output of a wavelet analysis stage can be considered as the
used techniques for image resolution enhancement is image to be interpolated. The original image can given as
Interpolation. Interpolation has been widely used in many input to a single wavelet synthesis stage along with the
image processing applications. Interpolation in image corresponding high frequency subbands to produce an
processing is a method to increase the number of pixels in a image interpolated by a factor of two in both directions. The
digital image. Traditionally there are three techniques for creation of unknown high-frequency subbands is necessary
image interpolation namely Linear, Nearest Neighbor and in the regularity-preserving interpolation strategy. Two-step
Bicubic. Nearest Neighbor result in significant Jaggy process is carried out to obtain the unknown high-frequency
edge distortion [1]. The Bilinear Interpolation result in subbands separably. In First step, in each row edges with
smoother edges but somewhat blurred appearance overall significant correlation across scales are identified.
[1]. Bicubic Interpolation looks best with smooth edges and Then near these edges the rate of decay of the wavelet
much less blurring than the bilinear result [1]. coefficients is extrapolated to approximate the high-
By applying the 1-D discrete wavelet transform (DWT) frequency subband required to resynthesize a row of twice
along the rows of the image rst, and then along the columns the original size. In second step, the same procedure as in
to produce 2-D decomposition of image[8]. DWT produce first step is then applied to each column of the row-
four subbands low-low(LL), low- high(LH), interpolated image. Block diagram of interpolation system
high-low(HL)and high-high(HH).By using these four for 1-D row and column signals is shown in Fig. 2.
subands we can regenerate original image[8]. Theoretically,
a lter bank shown in Fig. 1 should work on the image in
order to generate different subband frequency images.

Manuscript received October 15, 2013.

Sreeja K S, currently doing M-Tech in Electronics with specialization in

Signal Processing in Govt.College of Engineering Cherthala under Cochin
University of Science and Technology,Kerala,India
Prasanth C.R, currently pursing Mtech in Digital signal processing from
Govt. Engg college cherthala under Cochin University of science and
technology

68 www.erpublication.org
 Wavelet Transform Techniques for Image Resolution Enhancement

Fig 2. Block diagram of interpolation system for 1-D row and column resolution images are generated by spatial shifting, wavelet
signals[2]. transforming, and discarding the high frequency subbands.

ii)N high resolution images are obtained by applying the

III. NEW EDGE-DIRECTED INTERPOLATION WZP processing to all those low resolution images.

A hybrid approach produced by combining bilinear iii)The final high resolution image is reconstructed by re-
interpolation and covariance-based adaptive interpolation is aligning and averaging these intermediated high resolution
used in [3] to reduce the overall computational complexity. images. Fig. 3 shows the block diagram of the WZP- and
Traditional linear interpolation schemes (e.g., bilinear and
CS-based image super resolution[4].
bicubic) based on space-invariant models are not able to
capture the fast evolving statistics around edges and
consequently produce interpolated images with blurred edges
and annoying artifacts. Linear interpolation is good due to its
computational simplicity but not good due to its the
performance issue. Geometric regularity is very much
important for the visual quality of a natural image such as the
sharpness of edges and the freedom from artifacts. Without
the loss of generality, Xin Li & Michael T. Orchard
assume that the low-resolution image X i , j of size H W
directly comes from of size of 2H 2W, i.e. Y2i ,2 j X
i, j .
Fig3. Block Diagram of the WZP-and-CS-based image Resolution
They use the following basic problem to introduce their new
Enhancement [4].
interpolation technique: How do they interpolate the
interlacing lattice Y2i 1,2 j 1 from the lattice
Y2i ,2 j X i , j .They constrain their selves to the fourth- V. DT-CWT BASED IMAGE RESOLUTION
order linear interpolation. ENHANCEMENT
In this technique, as stated in [5],[8] dual-tree CWT
(DT-CWT) is used to decompose an input image into
different subband images. DT- CWT is used to decompose
The above equation is core part of this algorithm
an input low-resolution image into different subbands.
invented in [3]. In order to manage the computational
Then, the high-frequency subband images and the input
complexity, they used the following hybrid approach:
image are interpolated, followed by combining all these
covariance-based adaptive interpolation is only applied to
images to generate a new high-resolution image by using
edge pixels (pixels near an edge); for nonedge pixels
inverse DT-CWT. The resolution enhancement is achieved
(pixels in smooth regions), they still use simple bilinear
by using directional selectivity provided by the CWT,
interpolation. Such a hybrid approach is based on the
where the high-frequency subbands in six different
observation that only edge pixels benefit from the
directions contribute to the sharpness of the high-frequency
covariance-based adaptation and edge pixels often consist of
details, such as edges. Fig. 4 shows details of this
a small fraction of the whole image. A pixel is
technique, where the enlargement factor through the
considered as an edge pixel if an activity measure (e.g., the
resolution enhancement is .
local variance estimated from the nearest four neighbors) is
above a preselected threshold value . Since the computation
of the activity measure is typically negligible when
compared to that of covariance estimation, dramatic
reduction of complexity can be achieved for images
containing a small fraction of edge pixels. Xin Li &
Michael T. Orchard in [3] have found that the percentage of
edge pixels ranges from 5% to 15% for the test images used
in their experiments, which implies a speed-up factor of 7
20 [3].

IV. WZP-CS BASED IMAGE RESOLUTION Fig4. Block Diagram of DT-CWT Based Image Resolution
Enhancement [5]
ENHANCEMENT
As stated in[4],[8] this algorithm consists of two main steps The dual-tree complex wavelet transform (CWT) is
as follows: Step 1) an initial approximation to the unknown a relatively recent enhancement to the discrete wavelet
high resolution image is generated using wavelet domain zero transform (DWT), with important additional properties: It is
padding (WZP).Step 2) The cycle-spinning methodology is nearly shift invariant and directionally selective in two and
adopted to operate the following tasks: higher dimensions. It achieves this with a redundancy factor
i) Using high resolution image in part (1) a number of low of only 2d for d-dimensional

69 www.erpublication.org
 International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869, Volume-1, Issue-8, October 2013

signals, which is substantially lower than the undecimated directional resolution when compared with that of the
DWT. The multidimensional (M-D) dual-tree CWT is decimated DWT. Such features make it suitable for image
nonseparable but is based on a computationally efficient, resolution enhancement. In this letter, a complex
separable filter bank (FB).A method for image resolution wavelet-domain image resolution enhancement algorithm
enhancement from a single LR image using the dual-tree based on the estimation of wavelet coefficients at HR scales is
complex wavelet. The rough estimate of the HR image is proposed. The initial estimate of the HR image is constructed
decomposed to estimate the complex-valued high-pass by applying a cycle-spinning methodology in the DT-CWT
wavelet coefficients for the input LR image. The estimated domain. It is then decomposed using the one-level DT-CWT
complex wavelet coefficients are used, together with the input to create a set of high-pass coefficients at the same spatial
LR image, to reconstruct the resultant HR image by resolution of the LR image. The high-pass coefficients,
employing IDT-CWT.Image resolution enhancement is a together with the LR image, are used to reconstruct the HR
usable preprocess for many satellite image processing image using inverse DT-CWT (IDT-CWT).
applications. The DT-CWT is a combination of two real-valued
Image resolution enhancement is a usable preprocess decimated DWTs. The ordinary decimated DWT is shift
for many satellite image processing applications, such as variant due to the decimation operation exploited in the
vehicle recognition, bridge recognition, and building transform. As a result, a small shift in the input signal can
recognition to name a few. Image resolution enhancement result in a very different set of wavelet coefficients. For that,
techniques can be categorized into two major classes Kingsbury introduced a new kind of wavelet transform, called
according to the domain that they are applied in: 1) image the DT-CWT which exhibits shift-invariant property and
domain and 2) transform domain. The techniques in the image improves directional resolution when compared with that of
domain use the statistical and geometric data directly the decimated DWT.
extracted from the input image itself while transform-domain The DT-CWT also yields perfect reconstruction by using two
techniques use transformations such as decimated discrete parallel decimated trees with real-valued coefficients
wavelet transform (DWT) to achieve the image resolution generated at each tree. The 1-D DT-CWT decomposes the
enhancement. The decimated DWT has been widely used for input signal f(x) by expressing it in terms of a complex shifted
performing image resolution enhancement. A common and dilated mother wavelet (x) and a scaling function (x),
assumption of DWT-based image resolution enhancement is i.e.,
that the low-resolution (LR) image is the low-pass-filtered
subband of the wavelet-transformed high-resolution (HR)
image. This type of approach requires the estimation of
wavelet coefficients in sub bands containing high-pass spatial where Z is the set of natural numbers, j and l refer to the index
frequency information in order to estimate the HR image from of shifts and dilations, respectively, sj0,l is the scaling
the LR image. coefficient, and cj,l is the complex wavelet coefficient with
In order to estimate the high-pass spatial frequency j0,l(x) = rj 0,l(x) + 1i j0,l(x) and j,l(x) = r j,l(x)
information, many different approaches have been + 1i j,l(x), where the superscripts r and i denote the real
introduced.The high-pass coefficients with significant and imaginary parts, respectively. In the 1-D DT-CWT case,
magnitudes are estimated as the evolution of the wavelet the set {rj 0,l,i j0,l,rj 0,l,i j0,l
coefficients among the scales. The performance is mainly } forms a tight wavelet frame with double redundancy. The
affected from the fact that the signs of the estimated real and imaginary parts of the 1-D DT-CWT are computed
coefficients are copied directly from parent coefficients using separate filter banks with filters h0 and h1 for the real
without any attempt being made to estimate the actual signs. part and g0 and g1 for the imaginary part. Similar to the 1-D
This is contradictory to the fact that there is very little DT-CWT, the 2-D DT-CWT decomposes a 2-D image f(x, y)
correlation between the signs of the parent coefficients and through a series of dilations and translations of a complex
their descendants. As a result, the signs of the coefficients scaling function and six complex wavelet functions j,l,
estimated using extreme evolution techniques cannot be relied
upon. A hidden Markov tree (HMT)-based method models METHOD
the unknown wavelet coefficients as belonging to mixed Let us consider the unknown 2H 2W HR image
Gaussian distributions which are symmetrical about the zero XH and the known H W LR image XL. The aim of the
mean. HMT models are used to determine the most probable enhancement is to generate an estimated HR image XH of the
state for the coefficients to be estimated. The performance unknown HR image XH using the known LR image XL. Let
also suffers mainly from the sign changes between the scales. us further assume that the one-level DT-CWT decomposition
The decimated DWT is not shift invariant, and as a result, of a 2H 2W image X results in a matrix of DT-CWT(X) =
suppression of wavelet coefficients introduces artifacts into [LPX HPX], and the IDT-CWT of [LPX HPX] reconstructs
the image which manifest as ringing in the neighborhood of the signal X perfectly, i.e., IDT-CWT([LPX HPX]) = X. LPX
discontinuities. In order to combat this drawback in is a matrix of size H W which is the complex-valued
DWTbased image resolution enhancement, a cycle-spinning low-pass subband resulting
methodology was adopted. The perceptual and objective from the one-level DT-CWT decomposition of image X, and
quality of the resolution-enhanced images by their method HPX is a matrix of size H W 6 which is the collection of all
compares favorably with that in recent methods in the field. A six complex-valued high-pass subbands resulting from the
dual-tree complex wavelet transform (DT-CWT) is one-level DT-CWT decomposition of image X.
introduced to alleviate the drawbacks caused by the For a given LR image XL, the proposed resolution
decimated DWT . It is shift invariant and has improved enhancement method is made up of the following four main

70 www.erpublication.org
 Wavelet Transform Techniques for Image Resolution Enhancement

steps: 1) Generate the initial estimate (Y) of the HR image; 2) method is the same for all subband and the input images.
decompose Y using one-level DT-CWT to create a low and The interpolation technique and the wavelet function are
high-pass matrix structure [LPY HPY]; 3) formulate a matrix two important factors in determining the quality of the
structure [XL HPY] using [LPY HPY] and the input LR super-resolved images. To measure quality of image PSNR
image XL; and 4) generate the HR image by employing the value is used.
IDT-CWT on [XL HPY]. The first step employs the
cycle-spinning algorithm in the DT-CWT domain to create an
initial estimate of the unknown HR image.
The second step is the estimation of the high-pass
coefficients for the input LR signal XL. The initial estimate Y
is decomposed using the one-level DT-CWT to create one
complex valued low-pass sub band and six complex-valued
high-passsubbands with the same spatial resolution as that of
XL, i.e.,
DT-CWT(Y) = [LPY HPY].
In the final step, the input LR image, together with
the complex-valued high-pass sub bands HPY extracted from
the one-level DT-CWT decomposition of Y, is used to create
the HR image by employing IDT-CWT, i.e.,

The dual-tree CWT is a valuable enhancement of the

traditional real wavelet transform that is nearly shift invariant FIG 5. BLOCK DIAGRAM OF IMAGE SUPER RESOLUTION BASED
and, in higher dimensions, directionally selective. Since the ON INTERPOLATION OF WAVELET DOMAIN HIGH FREQUENCY
real and imaginary parts of the dual-tree CWT are, in fact, SUBBANDS AND THE SPATIAL DOMAIN INPUT IMAGE[6]
conventional real wavelet transforms, the CWT benefits from
the vast theoretical, practical, and computational resources
that have been developed for the standard DWT. A method VII. IMAGE RESOLUTION ENHANCEMENT METHOD
for image resolution enhancement from a single LR image USING SWT AND DWT [7]
using the dual-tree complex wavelet. The initial rough
estimate of the HR image is decomposed to estimate the The main loss in image resolution enhancement by using
complex-valued high-pass wavelet coefficients for the input interpolation is on its high frequency components (i.e.,
LR image. The estimated complex wavelet coefficients are edges), which is due to the smoothing caused by interpolation.
used, together with the input LR image, to reconstruct the Edges plays very important role in image. To increase the
resultant HR image by employing IDT-CWT.Thus image quality of the super resolved image, it is essential to preserve
resolution enhancement can be effectively achieved by using all the edges in image. In [7] work, DWT has been employed
dual-tree CWT. in order to preserve the high frequency components of the
image(i.e. edges). The redundancy and shift invariance of the
DWT mean that DWT coefficients are inherently
VI. IMAGE SUPER RESOLUTION BASED ON interpolable.In this correspondence, one level DWT (with
INTERPOLATION OF WAVELET DOMAIN HIGH Daubechies 9/7 as wavelet function) is used to decompose an
FREQUENCY SUBBANDS AND THE SPATIAL DOMAIN input image into different sub band images. Three high
INPUT IMAGE[6] frequency subbands (LH, HL, and HH) contain the high
frequency components of the input image(i.e. edges). In this
High-frequency components(i.e. the edges) are main loss
technique, bicubic interpolation with enlargement factor of 2
of an image after being super-resolved by applying
is applied to high frequency subband images. Information loss
interpolation. This loss occur due to the smoothing caused
occur due to downsampling in each of the DWT sub bands
by interpolation. To increase quality of the super-resolved
caused in the respective sub bands. That is why SWT
image, preserving the edges is essential. In [6] work by
(Stationary Wavelet Transform) is used to minimize this loss.
Hasan Demirel and Gholamreza Anbarjafari, DWT has
been employed in order to preserve the high-frequency The SWT is an inherently redundant scheme as the
components of the image. DASR technique uses DWT to output of each level of SWT contains the same number of
decompose an image into different subband images; samples as the input so for a decomposition of N levels
namely, low-low (LL), low-high (LH), high-low (HL), and there is a redundancy of N in the wavelet coefficients. The
high-high (HH). interpolated high frequency subbands and the SWT high
These subband images contain the high-frequency frequency subbands have the same size which means they
components of the input image. In the DASR technique, the can be added with each other. The new corrected high
interpolation is applied to high-frequency subband images. frequency subbands can be interpolated further for higher
This technique interpolates the input image as well as the enlargement. Also it is known that in the wavelet domain,
high-frequency subband images obtained through the DWT lowpass filtering of the high resolution image produce the
process. IDWT of the interpolated subband images and the low resolution image. In other words, low frequency
input image produce the final high-resolution output subband is the low resolution of the original image.
image. In the DASR technique, the employed interpolation Therefore, instead of using low frequency subband, which

71 www.erpublication.org
 International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869, Volume-1, Issue-8, October 2013

contains less information than the original high resolution

image, Hasan Demirel and Gholamreza Anbarjafari [7] are
using the input image for the interpolation of low frequency Where M and N are the size of the images. When the two
subband image. The quality of the super resolved image images are identical, the MSE will be zero.Clearly RMSE is
increases using input image instead of low frequency the square root of MSE, hence it is given by
subband. Fig. 6 illustrates the block diagram of the used
image resolution enhancement technique. RMSE=MSE
By interpolating input image by 3, and high frequency
subbands by 2 and in the intermediate and final And image entropy is a quantity which is used to describe
interpolation stages respectively, and then by applying the business of an image. , i.e. the amount of information
IDWT, as illustrated in Fig. 6, the output image will \which must be coded for by a compression algorithm.
contain sharper edges than the interpolated image obtained Image entropy is calculated with the formula
by interpolation of the input image directly. This is due to the
fact that, the interpolation of isolated high frequency ENTROPY=-PI Log2PI
components in high frequency subbands and using the
corrections obtained by adding high frequency subbands of In the above expression, P i is the probability that the
SWT of the input image, will preserve more high frequency difference between two adjacent pixels is equal to i, and
components after the interpolation than interpolating input Log 2 is the base 2 logarithms
image directly.

Fig 6. Block diagram of image resolution enhancement method using

SWT and DWT [7].

VIII. EXAMPLES AND DISCUSSION

These results are obtained by Hasan Demirel and
Gholamreza Anbarjafari as shown in table I from [7].
PSNR and Entropy values are used to measure the quality of
an image. Peak signal-to-noise ratio (PSNR) and root mean
square error (RMSE) have been implemented in order to TABLE I
obtain some quantitative results for comparison. PSNR (DB) VALUES FOR DIFFERENT RESOLUTION
ENHANCEMENT TECHNIQUES FROM 128x128 TO 512x5
PSNR can be obtained by using the following formula :

PSNR=10 log R2/MSE IX. CONCLUSION

Different image resolution enhancement techniques in
wavelet domain are discussed in this paper.WZP , is a simple
method we can approach for image resolution
enhancement.But efficiency is poor compared to other
Where R is the maximum fluctuation in the input image(255 methods. The DTCWT,DWT,SWT methods gives better
in here as the images are represented by 8 bit , i.e., 8- bit image resolution enhancement results.It is cleared from table I
grayscale representation have been used radiometric values that the image resolution enhancement method using
DWT & SWT is giving far better result than any other
resolution is 8 bit). technique studied in this paper.

APPLICATIONS
MSE=Iavg(i,j)-I(I,j)/(M*N) The image resolution enhancement techniques

72 www.erpublication.org
 Wavelet Transform Techniques for Image Resolution Enhancement

discussed in this paper can be apply to noisy images and

is mainly applicable to low resolution images.

REFERENCES
[1] Digital Image Processing Using MATLAB (Second Edition) by
Rafel C. Gonzalez and Richard E. woods and Steven L. Eddins.
[2] W. Knox. Carey, Daniel. B. Chuang, and S. S. Hemami, Regularity
Preserving image interpolation, IEEE Trans. Image Process., vol. 8, no. 9,
pp. 12951297, Sep. 1999.
[3] Xin. Li and Michael. T. Orchard, New edge-directed interpolation,
IEEE Trans. Image Process., vol. 10, no. 10, pp. 15211527, Oct. 2001.
[4] Alptekin. Temizel and Theo. Vlachos, Wavelet domain image
resolution enhancement using cycle-spinning, Electron. Lett., vol. 41, no.
3, pp. 119121, Feb. 3, 2005.
[5] Hasan. Demirel and Gholamreza. Anbarjafari, Satellite image resolution
enhancement using complex wavelet transform, IEEE Geosci. Remote
Sens. Lett., vol. 7, no. 1, pp. 123126, Jan. 2010.
[6] Gholamreza. Anbarjafari and Hasan. Demirel, Image super
resolution based on interpolation of wavelet domain high frequency
subbands and the spatial domain input image, ETRI J., vol. 32, no. 3, pp.
390394, Jun. 2010.
[7] Hasan. Demirel and Gholamreza. Anbarjafari, IMAGE Resolution
Enhancement by Using Discrete and Stationary Wavelet Decomposition,
IEEE IMAGE PROCESSING, VOL. 20, NO. 5, MAY 2011..
[8] Hasan. Demirel and Gholamreza. Anbarjafari, Discrete Wavelet
Transform-Based Satellite Image Resolution Enhancement, IEEE
Geosciences and Remote Sensing Letter, VOL. 49, NO. 6, JUNE
2011.Brown, L. D., Hua, H., and Gao, C. 2003. A widget
[9] Dual Tree Complex wavelet Transforms- Ivan W. Selesnick, Richard
G. Baraniuk, and Nick G. Kingsbury
[10] Image Resolution Enhancement using Dual Tree Complex Wavelet
Transforms- Turgay Celik and Tardi Tjahjadi

Sreeja K S received the B- tech degree from Rajagiri

School of Engineering and Technology under
Mahatma Gandhi University,Kochi,Kerala,India.She
is currently doing M-Tech in Electronics with
specialization in Signal Processing in Govt.College of
Engineering Cherthala under Cochin University of
Science and Technology,Kerala,India

Prasanth C.R Received Diploma in Electronics and

Instrumentation from govt. Polytechnic cherthala (first class with
distinction, third rank and gold medallist). He passed Btech in
Electronics and Communication Engg from M.G university kerala,
india.He currently pursing Mtech in Digital signal processing from Govt.
Engg college cherthala under Cochin University of science and
technology

73 www.erpublication.org