Cellular Automata for Image Resizing - Cellular Automata in Image Processing and Geometry - page 26

for learning statistical inter-pixel correlation of interpolated images. The method
in [21] comprises a hybrid artificial intelligence system. A fuzzy decision system is
proposed to classify all the pixels of the input image into human perception non-
sensitive class and sensitive class. The bilinear interpolation is applied to the non-
sensitive regions while a neural network is used to interpolate the sensitive regions
along the edges. Furthermore, the method proposed in [20] initially estimates local
covariance coefficients from a low resolution image. These covariance estimates are
used to adapt the interpolation at a higher resolution based on the geometric duality
between the low resolution and the high-resolution covariances. Despite the visually
accurate resulted images, the above edge directed approaches display high levels of
computational cost and thus, their application in real-time systems is restricted. In
order to achieve frame rates close to real time limits while enhancing the quality of
the edges, an edge-oriented method is proposed in [8]. The main idea is to seperate
the image homogenous areas and edges areas the latter processed using different
interpolation methods. The method achieves real-time image enlargement neverthe-
less, the classification of the areas depends on a predefined threshold as well as two
stages of process are required. Finally, Shi et al. in [33] initially expand the low
resolution image using a bilinear interpolation method and a Canny edge detector
is applied to identify the edges of the up-scaled image. The final light intensity val-
ues are calculated by applying some refinement functions. Despite the satisfactory
visual results and the high frame rates, the inaccuracies introduced by the initial
bilinear enlargement lead to blurred edges and thus, the Canny edge detector [4] is
unable to detect significant edges.
On the other hand, non conventional computational techniques like CA have been
applied to the field of image processing with great success. More specifically, CA
have been successfully applied to image processing due to their discrete, fully par-
allel with local interconnections nature. Taking into consideration that most of the
common image processing methods are characterized by high complexity and their
high requirements for memory and computational resources, the usage of CA in
image processing has intrigued the scientists a long time ago. More specifically,
Preston and Duff [28] demonstrated how major image processing tasks, such as im-
age segmentation, skeletonization, and filtering may be approached using the CA
methodology. Furthermore, specific applications of these image processing tasks in
both science and biomedicine are also presented. Lafe has also proposed CA meth-
ods, by which information building blocks, called basis functions (or bases), can
be generated from the evolving states, and called it Cellular Automata Transforms
(CAT) applying them to image and video compression in [19]. In recent years, it
has also been shown by several researchers [2, 12, 26, 30-32] that CA can be used
to perform some standard image processing tasks to a high level of performance,
as well as in up-to-date computer vision fields, such as stereo vision [9, 22, 24].
For example, Rosin in [30] proposed the training of CA to perform several im-
age processing tasks, namely noise filtering (also applied to grayscale images us-
ing threshold decomposition), thinning, and convex hulls, while the same author
proposed in [31] the application of CA in intensity images instead of binary ones,
able to perform many different image processing tasks, and that the quality of these