Neural Comput & Applic (****) **:** **
ORIGINAL ARTICLE
Scanned images resolution improvement using neural networks
Antigoni Panagiotopoulou Vassilis Anastassopoulos
Received: 1 August 2006 / Accepted: 1 February 2007 / Published online: 21 March 2007
Springer-Verlag London Limited 2007
Abstract A novel method of improving the spatial sequences (video) [3, 4]. In general, interpolation tech-
resolution of scanned images, by means of neural networks, niques produce a ne resolution image from a given LR
is presented in this paper. Images of different resolution, image. A great volume of reports on image interpolation
originating from scanner, successively train a neural net- using neural networks (NN) can be found in the literature
work, which learns to improve resolution from 25 to 50 [2, 3, 5 17]. The present work, which attempts to obtain a
pixels-per-inch (ppi), then from 100 to 200 ppi and nally, HR image by successively training the same NN, belongs
from 50 to 100 ppi. Thus, the network is provided with to this category. Neural networks describe systems without
consistent knowledge regarding the point spread function an explicit physical and mathematical modeling. Actually,
(PSF) of the scanner, whilst it gains the generalization the NN is trained to successfully carry out the solution of
ability to reconstruct ner resolution images unfamiliar to inverse problems, which in our case corresponds to
it. The novelty of the proposed image-resolution- restoring effects caused by scanner non-idealities.
enhancement technique lies in the successive training of Super-resolution techniques constitute a quite different
the neural structure with images of increasing resolution. image-processing tool. It is actually the estimation of a
Comparisons with the image scanned at 400 ppi dem- single HR image from multiple LR ones with subpixel
onstrate the superiority of our method to conventional shifts. The resulting image is much closer to the desired
interpolation techniques. one, compared to that obtained using interpolation tech-
niques [4]. Actually, frequency unfolding is achieved,
Resolution improvement Neural network which resolves details in the image ner than the sensor
Keywords
Scanner resolution limit. Neural networks have been used for
solving the super-resolution image reconstruction problem
[5, 6]. Nevertheless, it should be mentioned that from time
to time interpolation techniques are referred to as super-
1 Introduction
resolution, resulting in confusing naming.
Interpolation methods by means of neural networks have
Low-resolution (LR) images are commonly found in many
been employed for resolution enhancement in the area of
imaging applications, such as remote sensing, surveillance
remotely sensed images. An approach of spatial resolution
and astronomy. Over the last two decades, research has
improvement of a remotely sensed, LR, thermal infrared
been devoted to the problem of reconstructing a high-res-
image is described in [7]. The resolution improvement is
olution (HR) image from a LR frame. Basically, interpo-
carried out by means of two three-layered, backpropagated
lation techniques have been used to improve resolution in
NN architectures. Another approach of resolution
medical images [1], digital photos [2], as well as in image
improvement in remotely sensed images is that of a fully-
interconnected NN model [8]. The network used during
A. Panagiotopoulou V. Anastassopoulos testing is composed of three layers. Image interpolation has
Electronics Laboratory, Physics Department,
also been carried out by means of radial basis function
University of Patras, Rio 26500, Greece
neural networks in [9]. The hidden layer neurons of the
e-mail: abpihl@r.postjobfree.com
123
40 Neural Comput & Applic (2008) 17:39 47
image in order to construct the 400 ppi HR image. The
network which is employed to perform interpolation
proposed method outperforms the bicubic and spline
compute Gaussian basis functions. The receptive eld
algorithms, which belong to classical interpolation tech-
widths (standard deviations) of these functions are deter-
niques. The experimental results as well as the visual
mined by an adaptive scheme. Furthermore, HR image
comparisons prove the predominance of the proposed
reconstruction has been successfully carried out using
method. Indeed, the inversion of the scanner PSF and
multilayer perceptron networks in [10]. It should be noted
generalized image interpolation can be considered equiv-
that two different backpropagation algorithms are used,
alent. When performing interpolation at 200 ppi to obtain
whereas a multilayer perceptron network is also employed
the higher resolution image of 400 ppi, an effort towards
in [11]. The speci c single-hidden-layer neural network,
getting back missing samples of the image is made. Like-
being trained by the backpropagation algorithm, is required
wise, the NN, having as input the 200 ppi resolution image,
to enhance the resolution of diffraction-limited, binary
results in the 400 ppi image estimating the image samples
images. Moreover, a high-resolution, multi-neural network,
which were discarded during this image acquisition. Con-
based on the local variance is proposed in [12]. This spe-
sistent pieces of information regarding these samples can
ci c network is composed of two neural networks, namely
be obtained from the procedure applied to the HR image by
the NN for low local variance and the NN for high local
the scanner. The novelty of our approach lies in the pre-
variance. The weighted sum of the two NN outputs rep-
sentation order of image-pairs with different resolution
resents the enlarged image. A novel image interpolation
used to train the neural network. To the authors knowledge
scheme, using an arti cial neural network, is described in
there has not been any similar work in literature so far,
[13]. A single frame interpolation algorithm is joined to-
examining the importance of resolution order when pre-
gether with an adaptive, linear, single-layer neural network
senting a neural network with training data.
that models the residual errors between the interpolated
In Sect. 2 of this paper, the performance of the scanner
image and the respective original one. A novel image
as image source is presented. Section 3 consists of a de-
interpolation algorithm by means of a feedforward neural
tailed description of the training procedure, while Sect. 4
network, based upon classi cation, is thoroughly consid-
deals with the results of the conducted experiment. Sec-
ered in [14]. A HVS-oriented, adaptive interpolation
tion 5 presents some decisive aspects concerning our
scheme for natural images by means of neural networks is
treatment of the neural structure for the prementioned
proposed in [15]. Furthermore, a color interpolation tech-
purpose, whereas conclusions are drawn in Sect. 6.
nique for a single-chip CCD camera, employing neural
networks, is presented in [2]. A Hop eld-network-based
algorithm, serving for the resolution enhancement of dis-
crete targets taking up more space than the sample spacing 2 Using scanner as image source
of an image, is dealt with in [16]. Multilayer neural net-
works have also been used to perform document resizing in Scanning devices are able to produce images of the same
[17]. scene characterized by different resolution. The whole
In this paper a technique that performs image interpo- procedure that the scanner applies to an image is described
lation using neural networks is presented. A multilayer by
feedforward (MLFF) neural network, trained by the back-
n hx y io
propagation algorithm, is employed. The scanner provides xy
is n; m o x; y g x; y rect ; comb ;
all the input and output data that the network training re- ab xs ys
quires. Pairs of varying resolution images, obtained
1
through scanner, are used to successively train the network,
so that it becomes familiar with the scanner point spread where is (n, m) is the ensemble of the produced detector
function (PSF). More speci cally, the neural network is signals, o(x, y) is the object scene (input signal), g(x, y is
x y )
trained with the purpose of becoming able to reconstruct a the point spread function of the front optics and rect a ; b is
ner resolution image, even when its input is presented
the detector pupil. The 2D comb-function comb xxs ; yys
with images of different scenes from those than that it has
describes the scanning process. The symbol * denotes 2D
been trained for. Indeed, the resulting network improves
convolution. Scanner PSF is given by relation (1) if o(x, y)
the resolution of scanned images. Scanner, while recording
denotes a delta function. In accordance with the typical
an image, applies a kind of lowpass ltering to this original
de nition of scan directions, spatial scanning in the
high quality image and afterwards, by means of a speci c
direction of the detector line is called in-scan direction,
sampling, obtains the desired lower resolution image. In
while the orthogonal direction, which is along the direc-
this work the NN applies the inverse procedure of the
tion of scan motion for a 1D array, is given the name
scanner PSF, to a 200 pixels-per-inch (ppi) resolution
123
Neural Comput & Applic (2008) 17:39 47 41
cross-scan. Altogether it represents a 2D scanning process, followed, images of six different scenes are obtained from
which can be assumed to be separable and described in the scanner and subsequently used in the training and
both directions as a 1D process. During scanning with a simulation procedures. Changing the above mentioned
single 1D detector array, the spatial scanning frequency ks sampling distance p, each scene is recorded at 25, 50, 100,
is in reverse proportion to the middle distance p of 200 and 400 ppi resolutions.
neighboring detector pixels. This spatial sampling fre-
quency 1/p is equal to the distance of the delta impulses in
the comb-function [18]. During image acquisition, differ- 3 Training procedure
ent values can be assigned to the characteristic parameter p,
resulting in images of different resolution. Thus, the comb- A MLFF neural network is employed to perform the image
function is the element of (1), which takes various values resolution enhancement task. The speci c neural architec-
during the recording of different resolution images. ture is shown in Fig. 2. It is a hierarchical network con-
Moving from a high-resolution image to a lower-reso- sisting of an input layer, a hidden layer and an output layer.
lution one, a group of pixels is replaced by one pixel. In Each neuron of the input layer is connected through syn-
theory, this replacement is determined by the scanner optic weights to all the hidden layer neurons. The neurons
characteristics [18, 19]. A large pixel of LR image can be of the hidden layer are also connected to the four neurons
obtained from a group of HR image pixels, as shown in of the output layer in a similar way. Three layers are the-
Fig. 1, taking into consideration the scanner PSF and the 2D oretically suf cient to cope with any application. Usually,
sampling theorem. In case that the PSF of the scanner is the hidden layer possesses the greater number of neurons,
known, the unique evaluation of the LR image pixel from its as the contribution of these neurons to the successful
HR image counterparts can take place. Nevertheless, the execution of the desired mapping is considered crucial.
reverse procedure is not possible by means of a single LR Nevertheless, in the present case, experimentation proved
image, even if the scanner PSF is known. In our case, the
scanner PSF is considered unknown, thus the proposed
approach is universal, whereas the resulting neural network
succeeds in performing the inverse procedure of the scanner
PSF. This leads the network to enhancing the resolution of a
given image not having previously been trained upon it.
The reason for creating a neural network model with the
aforementioned ability lies in the scanner s capability of
providing us with varying resolution images. The use
of scanned images in this work provides the advantage of
having both the input and output vectors required to train
the network from the original source. Otherwise, LR images
have to be created from the HR data by means of decimation
procedures, an approach that demands knowledge of scan-
ner characteristics [18, 19]. In the experimental procedure
Fig. 2 The neural network architecture used at the proposed
interpolation approach. It is a fully interconnected structure, with
Fig. 1 The arrangement of pixels used to train the neural network. all outputs from the rst layer (left) going to all neurons of the second
The white pixels come from the low-resolution image whereas the (hidden) layer. In the same way, all outputs from the second (hidden)
black-shaded ones originate from the one-step-higher resolution layer are fed to all four neurons of the output (third) layer. Each
image. At the simulation procedure the aforementioned four pixels are neuron of the input layer is fed with the values of the 25 white pixels
created at the network output shown in Fig. 1
123
42 Neural Comput & Applic (2008) 17:39 47
that the rst two layers should have equal number of images, the neural network is supplied with consistent
neurons. These two layers consist of ten neurons, while the pieces of knowledge as far as the scanner PSF is concerned.
third one of four neurons, as required by the dimensions of In this way, it is successfully taught how to construct an
the output vector. The neural network performance has image characterized by resolution not previously presented
been also tested using, at the rst and second layers, the to it. The obtained resolution is higher than that of each of
following numbers of neurons: 5 and 10, 7 and 10, 10 and the images used as training data. Indeed, the resulting
15, 10 and 20 respectively, but the results were poorer neural network, when presented with a 200 ppi resolution
compared with those obtained using ten neurons. The image illustrating whatsoever scene, produces a satisfac-
backpropagation learning method is applied to the network. tory image of 400 ppi resolution.
Like the delta rule, it is an optimization procedure based on The neural network is not reset from one training stage
gradient descent that adjusts weights to reduce the system to another. Thanks to it, knowledge transfer regarding
error or cost function. The proposed training procedure of resolution improvement takes place from stage to stage.
the neural network is shown in Fig. 3. The training data The resolution order of the training data presentation to the
consist of four different resolution images, which illustrate network is another decisive point. The suggested presen-
the scene poros shown in Fig. 4. Speci cally, the ima- tation of the images results in an ef cient learning proce-
ges poros of 25, 50, 100 and 200 ppi resolutions are dure. Furthermore, the MSE obtained during training is of
order 10 2. Nevertheless, it is of crucial importance that the
employed to train the neural structure. These data are
successively fed to the network input and output. There is a network acquires satisfactory generalization ability. The
serial implementation of three training stages. It should be generalization ability is related to the successful recon-
stressed that each training stage deals with increasing res- struction of a 400 ppi resolution image, when the speci c
olution images of the scene poros . network is simulated with a 200 ppi resolution image. The
At the rst training step, the network is fed with the reconstruction should be successful not only in terms of the
25 ppi resolution image and the output consists of the image poros, as being the network input, but of various
50 ppi resolution image. Afterwards, the images with 100 scenes images as well. As it is shown below, the neural
and 200 ppi resolution are used to form the network input network obtains the desired generalization ability.
and output, respectively. At the nal, third step, the neural
network goes through training with the 50 ppi resolution
image as input and the 100 ppi resolution image as output. 4 Experimental results
At every one training step, the input image is given per
25-component vectors, while the output image is presented After the training procedure has been completed, the ob-
per 4-component vectors. These 25-element vectors are tained neural network goes through the simulation proce-
progressively selected from the LR image in such a way dure. The image poros is used for training the neural
that the whole image is covered. As shown in Fig. 1, at the network. During simulation, ve unfamiliar to the network
HR image reconstruction, the central pixel of each of these images, illustrating various scenes of resolution 200 ppi,
sets is replaced by four new pixels coming from the one- are successively presented to its input along with the image
step-higher resolution image. Each of the 4-element vectors poros of the same resolution. One of these new images,
that corresponds to the neural network output, is used to the image pylos, is shown in Fig. 5. Images of 400 ppi
form the high-resolution image. Thus progressively, all the resolution, which prove satisfactory when compared with
pixels of the low-resolution image are substituted and this the corresponding scanner-originated ones, are created at
image is transformed to the one-step-higher resolution one. the neural network output. Additionally, the trained neural
Due to the fact that every pair of images employed during structure is tested with noisy data and the results are in
each training stage is made up of increasing resolution favor of the proposed method.
Fig. 3 The proposed method of
creating a HR image is based on
the successive training of the
same neural network with
different resolution images
123
Neural Comput & Applic (2008) 17:39 47 43
Fig. 4 a The image poros of
400 ppi resolution coming from
the neural network. b The image
poros of 400 ppi resolution
obtained using the bicubic
algorithm. c The image
poros of 400 ppi resolution
obtained using the spline
algorithm. d The image
poros of 400 ppi resolution
coming from the scanner
4.1 Numerical and visual comparisons image obtained by the proposed method, approach the
original values more ef ciently compared with those of the
Besides visual comparison, our method is also assessed by images obtained by means of the two classical interpolation
means of mean square error (MSE), mean and standard methods. Visual comparison, Figs. 4 and 5, also proves the
deviation values in comparison with the bicubic and spline predominance of the proposed method. Comparison results
interpolation algorithms. Tables 1 and 3 report on these regarding the rest four images used during simulation have
estimated values as far as the image poros is concerned. shown to conclude in favor of the proposed method, too.
The proposed serial training procedure leads the neural From the above analysis it is obvious that the image
network to a 400 ppi resolution image that demonstrates resolution improvement carried out by means of our neural
MSE = 0.0405. This value is lower than the MSE = 0.0433 network outperforms the corresponding results of the spline
evaluated for the images resulting from the bicubic and and bicubic interpolation algorithms. It should be stressed
spline algorithms. Moreover, the mean value of the net- that the proposed method leads to a neural network, which
work-produced image is equal to 0.5263 and approximates is not specialized in improving the resolution of a speci c
satisfactorily the original image one, which is equal to scene image. The employed neural structure, through the
0.5340. Images coming from the bicubic and spline presented training procedure, learns how to improve image
methods display mean values 0.5200 and 0.5034, respec- resolution and thus successfully constructs HR images not
tively, which are quite distant from the mean value of the having previously been trained upon. All the MSEs are
scanner-originated image. Furthermore, the standard devi- evaluated using the 400 ppi resolution image coming from
ation value, equal to 0.2329, of the network-produced the scanner. It is noted that the range of the pixel values is
image is quite close to the desired value of 0.2615 in 0 1, whereas the programming tool used is MATLAB.
contradiction with the values 0.2103 and 0.1921 of the
bicubic and spline algorithms images. 4.2 Edge details assessment
Additionally, the obtained neural network demonstrates
satisfactory results when it is simulated with images that In order to provide a more detailed assessment of the
have not been used during its training procedure. Tables 2 proposed method a small part of the image poros,
and 3 contain numerical results for the image pylos . The shown in Fig. 6, is acquired by means of the scanner at 25,
0.0409 MSE of our method is quite lower than the 0.0428 50, 100, 200 and 400 ppi resolutions. This speci c scene
and 0.0424 errors of the bicubic and spline interpolations. image is then used to train the neural network of Fig. 2. Its
Furthermore, the mean and standard deviation values of the simulation with the 200 ppi resolution image leads to the
123
44 Neural Comput & Applic (2008) 17:39 47
Fig. 5 a The image pylos of
400 ppi resolution coming from
the neural network. b The image
pylos of 400 ppi resolution
obtained using the bicubic
algorithm. c The image
pylos of 400 ppi resolution
obtained using the spline
algorithm. d The image
pylos of 400 ppi resolution
coming from the scanner
Table 1 Results in terms of mean and standard deviation values for Table 2 Results in terms of mean and standard deviation values for
the image poros of 400 ppi resolution the image pylos of 400 ppi resolution
Image origin Mean value Standard deviation value Image origin Mean value Standard deviation value
Network output 0.5263 0.2329 Network output 0.4773 0.1743
Bicubic 0.5200 0.2103 Bicubic 0.4976 0.1684
Spline 0.5034 0.1921 Spline 0.5066 0.1594
Scanner 0.5340 0.2615 Scanner 0.4844 0.2207
image of 400 ppi resolution illustrated in Fig. 6a. The avoiding the introduction of even small artifacts in com-
corresponding result obtained by means of the bicubic parison with the bicubic algorithm. The above result is
method is shown in Fig. 6b. Visual comparison proves the quanti ed via the edge correspondence measure. Prewitt
predominance of the proposed method as far as edge pre- edge detector is used to nd edges in the 400 ppi image of
serving is concerned. Our neural network achieves detailed Fig. 6a, in the corresponding image coming from the bic-
and exact reconstruction, while the conventional interpo- ubic method and in the original small part of the image
lation technique is not satisfactorily edge preserving. The poros originating from the scanner. Then, the number of
outline of the small window cited just above the door next correct edges that the images of Fig. 6a and b present is
to the white car and the right side of the black door coming calculated. The image of the proposed method has 68
immediately after the dark coloured car indicate that the correct edges, while the image obtained by means of the
proposed method is more capable of retaining edges and bicubic algorithm presents 44 correct edges. The spline
123
Neural Comput & Applic (2008) 17:39 47 45
transfer function. The Powell-Beale version of the conju-
Table 3 Results in terms of MSE for the images poros and
pylos of 400 ppi resolution gate gradient algorithm is the selected training function.
Furthermore, the training parameter concerning the mag-
Real output origin MSE between real output and target output
nitude of the gradient is set to the value 10 5, so that it can
Poros Pylos
be never reached. As far as the learning parameter is
concerned, it is set equal to 10 2. The number of training
Network output 0.0405 0.0409
epochs for each stage is 250. Each output image recon-
Bicubic 0.0433 0.0428
struction requires the estimation of (n1 4) (n2 4)
Spline 0.0433 0.0424
4-component vectors, where n1, n2 denote the number of
the input image pixels in the vertical and horizontal
directions, respectively. The calculation of each of the
above 4-component vectors is performed by a neural
network that functions having 25 10 + 10 10 +
10 4 = 390 weights. The number of coef cients used by
the bicubic interpolation algorithm is 13, while the spline
interpolation uses 361 coef cients. Thus, the computa-
tional complexity of our neural-network-based method is
of the same order as that of the spline interpolation, but of
higher order than that of the bicubic interpolation.
4.5 Remarks
Fig. 6 a A small part of the image poros of 400 ppi resolution
coming from the neural network. b A small part of the image poros
It should be mentioned that during the training procedure
of 400 ppi resolution obtained using the bicubic algorithm. It is
the neural network may reach a speci c value of training
obvious that the proposed method outperforms the bicubic interpo-
MSE but cannot proceed to a lower level. If it is forced
lation as far as edge preserving is concerned
by the number of training epochs to repeatedly reach this
MSE value, due to continuous training, it becomes un-
algorithm outcome is not presented here but is the same as able to ful ll our expectations concerning generalization
that of the bicubic algorithm. ability. More clearly, overtraining occurs because the
neurons memorize the training set patterns and become
4.3 Performance under noisy conditions unable to recognize target-class patterns that were not
included in the training set. Thus, using a number of
Furthermore, the performance of the obtained neural net- epochs larger than the 250 selected is not recommended.
work under noisy conditions is examined. Gaussian noise In contrast, fewer than 250 training epochs are inade-
with variance 25 is added to the image poros of 200 ppi quate for ef cient learning of the neural network.
resolution, which is then fed to the network input. This Overtraining also takes place in case of using more than
degraded image is shown in Fig. 7a. The 400 ppi resolution ten neurons in the rst two layers, whereas fewer than
image produced at the network output is depicted in ten neurons are insuf cient in view of our purpose.
Fig. 7b. Figure 7c and d illustrate the result of creating a Additionally, when choosing a value of the learning rate
parameter higher than 10 2, the whole procedure may not
400 ppi resolution image from the degraded 200 ppi res-
olution image, by means of the bicubic and spline inter- converge to the appropriate network weight values [20].
polation algorithms. It is visually obvious that the proposed Numerous combinations of activation functions used by
method outperforms the aforementioned two conventional the neural network neurons have also been tested. This
interpolation techniques. The neural network results in an aforementioned combination tansig tansig logsig was the
image with noise almost eliminated, in contradiction with most ef cient. Moreover, the presentation order of
the images coming from the bicubic and spline interpola- training data to the neural network has proved to be of
tion algorithms. great signi cance, as well. The serial training procedure
has also been tested, with input output image pairs
4.4 Computational details 25 50, 50 100 and 100 200 ppi. However, the obtained
network appears to possess limited abilities. Furthermore,
The neurons of the rst two layers use the tan-sigmoid various combinations of the prementioned three training
transfer function to generate their outputs, while the acti- stages for the implementation of a nine-stages-repetitive
vation function of the last layer neurons is the log-sigmoid training procedure have been examined. Nevertheless,
123
46 Neural Comput & Applic (2008) 17:39 47
Fig. 7 a The image poros of
200 ppi resolution degraded by
Gaussian noise. b The image
poros of 400 ppi resolution
coming from the neural network
when it is presented with the
degraded 200 ppi resolution
image. c The image poros of
400 ppi resolution obtained
when performing bicubic
interpolation at the degraded
200 ppi resolution image. d The
image poros of 400 ppi
resolution obtained when
performing spline interpolation
at the degraded 200 ppi
resolution image
this repetitive implementation did not prove ef cient as that neural structures cannot easily retain the knowledge
the resulting image was over-smoothed. gained by means of a speci c set of training data when a
new training data set is presented to them. Furthermore,
during the training procedure the network employed for
interpolation is provided with consistent knowledge of the
5 Discussion
standard deviation and mean values of the image of ner
resolution. In reverse, when creating an interpolated image
The fact that the neural network weights are not reinitial-
through the bicubic and spline methods, there are no
ized from training step to training step on the one hand, and
a priori pieces of knowledge of these characteristic values
the order the training images are fed to the network on the
of the constructed image.
other, are two decisive points of the presented strategy.
Both result in retaining the neural network knowledge from
one training stage to another. An additional signi cant
point is the neural structure training with images coming 6 Conclusions
directly from the scanner. Due to the originality of this
training procedure, the network is provided with the In this work, a multilayer feedforward neural network is
appropriate pieces of knowledge, so that it can learn to employed to improve the resolution of scanned images.
improve image resolution. Moreover, when treating a Pairs of different resolution images, illustrating the scene
neural network special attention should be paid to avoid poros and coming from the scanner, are used in the
overtraining. The number of the training epochs as well as training procedure. There are three training stages each of
the number of the neurons in each layer must be carefully which deals with a pair of increasing resolution images. The
selected. It should be stressed that the low value of the neural network gains knowledge concerning resolution
MSE obtained during training does not keep pace with a enhancement from 25 to 50 ppi, from 100 to 200 ppi and
visually satisfactory image of 400 ppi derived from the from 50 to 100 ppi. In this way it achieves to ef ciently
network, as it differs from the adequate generalization learn the PSF of the scanner. The rationale behind the
ability. This ability is indeed connected to the desired, proposed approach is the training of the network in order
successful reconstruction of the HR image. Nonetheless, it not only to learn how to improve a speci c image resolution
should be stated that the visual inspection remains the most to a ner one, but to generalize this knowledge to even ner
reliable measure of comparison. Moreover, it is evident resolution and different scene image, as well. Actually, the
123
Neural Comput & Applic (2008) 17:39 47 47
7. Valdes M del C, Inamura M (1998) Spatial resolution improve-
suggested successive presentation of the different resolution
ment of a low spatial resolution thermal infrared image by
training data to the network has led to the successful
backpropagated neural networks. IEICE Trans Inf Syst E81
application of the inverse procedure of the scanner PSF to D:872 880
200 ppi resolution images that are unfamiliar to it. Satis- 8. Valdes M del C, Inamura M (2000) Spatial resolution improve-
ment of remotely sensed images by a fully interconnected neural
factory images of 400 ppi resolution illustrating even dif-
network approach. IEEE Trans Geosc Rem Sens 38:2426 2430
ferent scenes from the scene poros are the outcome of
9. Ahmed F, Gustafson SC, Karim MA (1996) Image interpolation
this attempt. Visual comparisons, MSE, mean and standard with adaptive receptive eld-based Gaussian radial basis func-
deviation values have proved that the proposed method tions. Microw Opt Techn Lett 13:197 202
10. Plaziac N (1999) Image interpolation using neural networks.
predominates over the classical bicubic and spline inter-
IEEE Trans Image Process 8:1647 1651
polation algorithms. This method reconstructs images of
11. Davila CA, Hunt BR (2000) Superresolution of binary images
400 ppi resolution that successfully approach the corre- with a nonlinear interpolative neural network. Appl Opt 39:2291
sponding original images obtained by means of the scanner. 2299
12. Sekiwa D, Taguchi A (2001) Enlargement of digital images by
The proposed method also performs superiorly even under
using multi-neural networks. Electr Commun Jpn 84:61 69
noisy conditions. The present method is novel as it treats,
13. Pan F, Zhang L (2003) New image super-resolution scheme based
for the rst time, the issue of the training data presentation on residual error restoration by neural networks. Opt Eng
order to the neural network input. 42:3038 3046
14. Hu H, Hofman PM, Haan G (2004) Image interpolation using
classi cation-based neural networks. IEEE Intern Symp Consum
Electr 1:133 137
References
15. Pu H-C, Lin C-T, Liang S-F, Kumar N (2003) A novel neural-
network-based image resolution enhancement. IEEE Intern Conf
1. Lehmann TM, Gonner C, Spitzer K (1999) Survey: interpolation Fuzz Syst 2:1428 1433
methods in medical image processing. IEEE Trans Med Imaging 16. Collins M, Jong M (2004) Neuralizing target superresolution
18:1049 1075 algorithms. IEEE Geosc Rem Sens Lett 1:318 321
2. Go J, Sohn K, Lee Ch (2000) Interpolation using neural networks 17. Ahmed MN, Cooper BE, Love ST (2001) Document resizing
for digital still cameras. IEEE Trans Consum Electr 46:610 616 using a multi-layer neural network. NIP17: Intern Conf Dig Print
3. Skoneczny S, Szostakowski J (2001) Classical and neural meth- Techn:792 796
ods of image sequence interpolation. Proc SPIE 4535:191 204 18. Craubner S (2002) Optoelectronic image scanning with high
4. Tekalp AM (1995) Digital video processing. Prentice-Hall, USA spatial resolution and reconstruction delity. Opt Eng 41:381
5. Salari E, Zhang S (2003) Integrated recurrent neural network for 394
image resolution enhancement from multiple image frames. IEE 19. Shen H-L, Xin JH (2004) Spectral characterization of a color
Proc 150:299 305 scanner by adaptive estimation. J Opt Soc Am 21:1125 1130
6. Tsagaris V, Panagiotopoulou A, Anastassopoulos V (2004) 20. Patterson DW (1996) Arti cial neural networks: theory and
Interpolation in multispectral data using neural networks. Proc applications. Prentice-Hall, Singapore
SPIE 5573:460 470
123