Information Technology Reference
In-Depth Information
texture synthesis methods, the works of Efros and Leung [1] and Wei and Levoy [2]
have been influential. Both methods are pixel-based, and Wei and Levoy's method
is basically an improved and more efficient version of [1]. The algorithm
synthesizes an image in a raster scan order and for each unfilled pixel it does a full
search in the sample image to find a pixel whose neighbourhood matches the best to
the filled neighbourhood of the unfilled pixel. Tree-structured Vector Quantisation
is used to significantly speed up the search process. Ashikhmin [3] presented a
constrained search scheme in which only shifted locations of neighbourhood pixels
of the current unfilled pixel are used for matching. Patch-based texture synthesis is
proposed by Efros and Freeman [4] and Kwatra et al. [5]. These techniques are
usually more effective for structured textures and are faster than pixel-based
methods.
The above mentioned synthesis methods do not consider the structural layout in
the input sample. Therefore, the output image would look very different to the input
globally though is composed of the same textures. For the purpose of image scaling,
the proposed algorithm can capture both the texture elements and the global layout
of the input. For an image composed entirely of texture, we use a target image
which has the same size as the output to guide the synthesis process. For an image
composed of both a foreground object and the background texture, foreground
object is segmented and treated differently to the background texture.
In the remainder of this chapter, we will first describe in Section 2 the Image
Quilting algorithm introduced by Efros and Freeman [4] and use it as a benchmark for
our proposed method. In Section 3, a novel synthesis algorithm particularly for
natural composite textures is introduced and compared with state-of-the-art
techniques. The texture synthesis technique based on foreground object segmentation
is presented in Section 4. Finally, we conclude this chapter in Section 5.
2 Image Quilting
The Image Quilting (IQ) algorithm [4] is an efficient method to synthesize textures,
taking the place of pixel based algorithms which are slow and unreliable. The IQ
algorithm finds the best matched patches within certain error tolerance, called
candidate blocks
, from the input texture, and randomly selects one of the candidates
to stitch with the existing patches of the output image. The
stitching
is implemented
by finding the minimum error boundary cut in the overlap region. In this case, the
IQ algorithm guarantees the smoothness of the output image. Generally, the image
quilting algorithm works as follows:
•
Go through the image to be synthesized in the raster scan order in steps of one
block (minus the overlap).
•
For every location, search the input texture for a set of blocks that satisfy the
overlap constraints (above and left, see Fig. 1) within certain error tolerance.
Randomly pick one such block.
