MultiScale Photomosaic

1
Multi-Scale Photomosaic
Kihwan Kim and Irfan Essa Georgia Institute of
Technology
Overall process
Result Images
Introduction
To make the multi-scale mosaic, we first segment
the images and analyzing gradient data. Once we
make the gradient map and segmented look up area,
we can label multi-scaled tiles. As for Tile
image set which will be components of the mosaic
image, they are trained by obtaining properties
as follows
Goal
Making mosaic image tiling with multi-scaled
photo images
Abstract

• Average and Variance of the color information
• Previously Annotated data
• Histogram orientation data in every quadrant.

We introduce a novel method to tile various size
of photo images which represent annotated meaning
and color of the source image. We also suggest
efficient way to make multi-scaled tile
distribution in segmented area. By this method,
we can make the mosaic image without modification
of tiling photo images.
After Labeling and Training step, Matching
algorithm will allocate and categorize most close
tile images into previously made multi-scaled
tile area.. Matching step also include some after
effect to blend the boundary of each joint tile
images.
Labeling Tile object
Integral image represent the summation of every
density value within certain size of patch
(tile). If the Integral Image is equal to
,we can consider the tile as Labeled Tile
is uniformly distributed value of a pixel
point, is uniformly distributed value of
a pixel point in segmented area, is
uniformly distributed value of a pixel point in
segmented area where previously checked by other
Labeled tile, is uniformly distributed
value of a pixel point in background area which
is set to zero.
Matching algorithm
ltLabeling example 1gt
T1,T2,T3 is candidate tile set
Matching algorithm consists of 4 steps (1)
Calculating average color and variance distance
(2) Matching the boundary images using
Orientation Histogram (3) Match the annotated
information (4) After effect for boundary area
between interconnected tiles. Selecting most
close tile is calculated by following criteria.
T1
T2
Segmented Area

T3
Thus, T3 is chosen as Labeled tile.
So, every in T3 area is set to
.
is a pattern index from Histogram
Orientation analysis, is annotation
index. is minimum distance between image
information of input image in certain tiled area
and every trained images. However, if there is no
annotation information in source image and
training images, is determined only by
and . . Likewise, if we use an
input image as logo or letter image which has no
orientation information except boundary, is
determined by other factors. Thus, in this paper,
most of test was fulfilled without orientation
information.
ltLabeling example 2gt
After T3 is chosen, new candidate T4 comes
Conclusion

T3

T3
By using the multi-scaled tiling algorithm, we
made relatively desirable output images.
Furthermore, using an integral image method
boosts up the speed for calculating labeling
process. However, there are still some problems
to solve. First, we should make more appropriate
matching algorithm to compose mosaic images.
Second, when the training image set is not
sufficient to represent some color or some
meanings, the output will be very coarse.
Finally, because we used unmodified images as
tile images, interconnected boundary seems
somewhat unnatural. Thus, we also should apply
more proper method to solve these problems like
additional annotation matching structure,
blending and Cut off algorithm for boundary
modification and previously made Quadrant based
Orientation Histogram to the non-symbol image (
using for background to symbol boundary
modification). Furthermore, as a practical
application of these mosaicking method, we are
considering Media browser and Headline image
categorizer.
T4
T4
Because T4 cross the area where T3 already
occupied, some pixels in T4 have distribution
value as . So, the Integral image of T4 will
not meet .
Typomosaic
Typography is a design category of letters in
various fields. In Typomosaic sense, the letter
or phrase consists of the certain meaningful
image. And the subset of image (or video clips)
will represent the metaphor of the letter or
previously annotated information. When we see the
Typomosaic, we can see the many kind of (various
size) images which represent the some metaphor of
the letters and we can easily select the certain
images and see the original image of tiled one.