IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. XX,NO. XX, XXX. XXXX 3
(a) original (b) IT [17] (c) AIM [22] (d) IM [23] (e) MSS [24] (f) SEG [25] (g) SeR [26] (h) SUN [27] (i) SWD [28]
(j) GB [29] (k) SR [30] (l) AC [31] (m) CA [32] (n) FT [33] (o) LC [34] (p) CB [35] (q) HC (r) RC
Fig. 2. Saliency maps computed by different state-of-the-art methods (b-p), and with our proposed HC (q) and RC
methods (r). Most results highlight edges, or are of low resolution. See also Fig. 9 and our project webpage.
conspicuous parts and permit combination with other
importance maps. The model is simple, biologically
plausible, and easy to parallelize. Liu et al. [2] find
multi-scale contrast by linearly combining contrast in
a Gaussian image pyramid. More recently, Goferman
et al. [32] simultaneously model local low-level clues,
global considerations, visual organization rules, and
high-level features to highlight salient objects along with
their contexts. Such methods using local contrast tend
to produce higher saliency values near edges instead
of uniformly highlighting salient objects (see Fig. 2).
Note that Reinagel et al. [37] observe that humans tend
to focus attention in image regions with high spatial
contrast and local variance in pixel correlation.
Global contrast based methods evaluate saliency of an
image region using its contrast with respect to the entire
image. Zhai and Shah [34] define pixel-level saliency
based on a pixel’s contrast to all other pixels. How-
ever, for efficiency they use only luminance information,
thus ignoring distinctiveness clues in other channels.
Achanta et al. [33] propose a frequency tuned method
that directly defines pixel saliency using a pixel’s color
difference from the average image color. The elegant ap-
proach, however, only considers first order average color,
which can be insufficient to analyze complex variations
common in natural images. In Figures 9 and 12, we
show qualitative and quantitative weaknesses of such
approaches. Furthermore, these methods ignore spatial
relationships across image parts, which can be critical
for reliable and coherent saliency detection (see Sec. 6).
Saliency maps are widely employed for unsupervised
object segmentation: Ma and Zhang [41] find rectangular
salient regions by fuzzy region growing on their saliency
maps. Ko and Nam [43] select salient regions using
a support vector machine trained on image segment
features, and then cluster these regions to extract salient
objects. Han et al. [44] model color, texture, and edge
features in a Markov random field framework to grow
salient object regions from seed values in the saliency
maps. More recently, Achanta et al. [33] average saliency
values within image segments produced by mean-shift
segmentation, and then find salient objects via adaptive
thresholding. We propose a different approach that ex-
tends GrabCut [38] method and automatically initialize
it using our saliency detection methods. Experiments on
our 10, 000 images dataset (see Sec. 6.1.1) demonstrate
the significant advantages of our method compared to
other state-of-the-art methods.
Subsequent to our preliminary results [1], Jiang et
al. [35] propose a comparable method also making use of
region level contrast to model image saliency. In the seg-
mentation step, their method also expands and shrinks
the initial trimap and iteratively applies graphcut and
histogram appearance model. Since GrabCut is an iter-
ative process of using graphcut and GMM appearance
mode, the two segmentation methods share a strong sim-
ilarity. Compared to the CB method [35], experimental
results show that our RC salient object region detection
and segmentation is more accurate (Fig. 12(a)(c)), 20×
faster (Fig. 7), and more robust to center-bias (Fig. 12(b)).
3 HISTOGRAM BASED CONTRAST
Our biological vision system is highly sensitive to con-
trast in visual signal. Based on this observation, we pro-
pose a histogram-based contrast (HC) method to define
saliency values for image pixels using color statistics of
the input image. Specifically, the saliency of a pixel is
defined using its color contrast to all other pixels in the
image, i.e., the saliency value of a pixel I
k
in image I is,
S(I
k
) =
X
∀I
i
∈I
D(I
k
, I
i
), (1)
where D(I
k
, I
i
) is the color distance metric between pix-
els I
k
and I
i
in the L
∗
a
∗
b
∗
space for perceptual accuracy.
(1) can be expanded by pixel order as,
S(I
k
) = D(I
k
, I
1
) + D(I
k
, I
2
) + · · · + D(I
k
, I
N
), (2)
where N is the number of pixels in image I. It is
easy to see that pixels with the same color have the
same saliency under this definition, since the measure is
oblivious to spatial relations. Thus, rearranging (2) such
that the terms with the same color value c
j
are grouped
together, we get saliency value for each color as,
S(I
k
) = S(c
l
) =
X
n
j=1
f
j
D(c
l
, c
j
), (3)
剩余13页未读，继续阅读
等待不系之舟
- 粉丝: 5
- 资源: 10
我的内容管理
展开
最新资源
- 李兴华Java基础教程：从入门到精通
- U盘与硬盘启动安装教程：从菜鸟到专家
- C++面试宝典：动态内存管理与继承解析
- C++ STL源码深度解析：专家级剖析与关键技术
- C/C++调用DOS命令实战指南
- 神经网络补偿的多传感器航迹融合技术
- GIS中的大地坐标系与椭球体解析
- 海思Hi3515 H.264编解码处理器用户手册
- Oracle基础练习题与解答
- 谷歌地球3D建筑筛选新流程详解
- CFO与CIO携手：数据管理与企业增值的战略
- Eclipse IDE基础教程：从入门到精通
- Shell脚本专家宝典：全面学习与资源指南
- Tomcat安装指南：附带JDK配置步骤
- NA3003A电子水准仪数据格式解析与转换研究
- 自动化专业英语词汇精华：必备术语集锦
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
