700 Lin et al. / J Zhejiang Univ-Sci C (Comput & Electron) 2014 15(9):697-716
Fig. 3 Saliency based on different approaches (image adapted from Goferman et al. (2012)): (a) input image;
(b) Walther and Koch (2006)’s method; (c) Hou and Zhang (2007)’s method; (d) Liu et al. (2011)’s method;
(e) Goferman et al. (2012)’s method
tion to more applications handling different natural
images.
Because importance measures are subjective,
users can interactively specify the important regions.
Santella et al. (2006) minimized information about
the location of important content provided by eye
tracking. They used fixation data to identify impor-
tant content. Users were requested to mark those
parts on the image whose shapes should be preserved
(Gal et al., 2006).
3 Image resizing techniques
Image resizing techniques mainly include scal-
ing, cropping, seam carving, warping, multi-
operator, and other methods. Intelligent cropping,
seam carving, warping, and multi-operator resizing
are corresponding to content-aware resizing meth-
ods. In this section, we briefly describe the basic
theories of these methods and some new methods in
recent years.
3.1 Scaling
Scaling is defined by a homogeneous map be-
tween pixels of the original image and pixels of the
target image. The most common approach for scal-
ing adopts the interpolation of original image pix-
els. Nearest neighbor interpolation, bilinear interpo-
lation, and bi-cubic interpolation are the three most
commonly used interpolation methods. Image scal-
ing can be performed in real time and the global
visual effects can be preserved when interpolation
methods are employed. However, these interpola-
tion scaling methods can bring artifacts, such as an
artificial block and aliasing. Scaling causes obvious
distortion if the aspect ratio of the input image is
obviously different from that of the output image.
3.2 Cropping
The cropping method extracts a rectangular
window with a desired size from the original image.
The content within the window is kept and others are
discarded. The traditional cropping method simply
crops a cropping rectangle from the center of the
image as its output resizing result. This method is
very simple. However, it has a limitation of losing
those important contents lying on the periphery of
an image. So, the effect is seriously damaged.
To preserve important contents of an image,
a user can directly draw a crop rectangle around
them. However, it is time-consuming and burden-
some. So, content-aware intelligent cropping is pro-
posed to solve the problem. The intelligent crop-
ping method usually includes two steps, i.e., main
content detection and cropping. Suh et al. (2003)
used a saliency map to express important informa-
tion. They also considered the semantic information
such as face detection to enhance the result of auto-
matic thumbnail cropping. This method is substan-
tially more recognizable than the original cropping
method, but it depends on the detection algorithm
which often produces inaccurate results. Chen et al.
(2003) introduced an image attention model which
has three attributes, i.e., region of interest (ROI),
attention value (AV), and minimal perceptible size
(MPS). An attention object (AO) often represents a
semantic object, such as a human face and a text
sentence. Three attention models (saliency, face,
and text) were used to calculate their attention val-
ues respectively. They classified an image into five
different categories. Then different rules were used
to adjust the AV weights for different classes. Fi-
nally, a branch-and-bound algorithm was developed
to find the optimal adaptation efficiently. However,
this method heavily relies on semantic extraction
techniques. If the corresponding semantic technique