没有合适的资源?快使用搜索试试~ 我知道了~
资源推荐
资源详情
资源评论
Salient Object Detection: A Discriminative Regional Feature
Integration Approach
Huaizu Jiang
†
Jingdong Wang
‡
Zejian Yuan
†
Yang Wu
§
Nanning Zheng
†
Shipeng Li
‡
†
Xi’an Jiaotong University
‡
Microsoft Research Asia
§
Kyoto University
https://sites.google.com/site/jianghz88/saliency_drfi
Abstract
Salient object detection has been attracting a lot of
interest, and recently various heuristic computational
models have been designed. In this paper, we regard
saliency map computation as a regression problem. Our
method, which is based on multi-level image segmenta-
tion, uses the supervised learning approach to map the
regional feature vector to a saliency score, and finally
fuses the saliency scores across multiple levels, yielding
the saliency map. The contributions lie in two-fold.
One is that we show our approach, which integrates
the regional contrast, regional property and regional
backgroundness descriptors together to form the master
saliency map, is able to produce superior saliency maps
to existing algorithms most of which combine saliency
maps heuristically computed from different types of fea-
tures. The other is that we introduce a new regional fea-
ture vector, backgroundness, to characterize the back-
ground, which can be regarded as a counterpart of the
objectness descriptor [2]. The performance evaluation
on several popular benchmark data sets validates that
our approach outperforms existing state-of-the-arts.
1. Introduction
Visual saliency has been a fundamental problem in
neuroscience, psychology, neural systems, and com-
puter vision for a long time. It is originally a task of
predicting the eye-fixations on images, and recently has
been extended to identifying a region containing the
salient object, which is the focus of this paper. There
are various applications for salient object detection, in-
cluding object detection and recognition [25, 46], im-
age compression [21], image cropping [35], photo col-
lage [17, 47], dominant color detection [51, 52] and so
on.
The study on human visual systems suggests that
the saliency is related to uniqueness, rarity and sur-
prise of a scene, characterized by primitive features
like color, texture, shape, etc. Recently a lot of efforts
have been made to design various heuristic algorithms
to compute the saliency [1, 6, 11, 15, 18, 27, 31, 34, 38].
In this paper, we regard saliency estimation as a
regression problem, and learn a regressor that directly
maps the regional feature vector to a saliency score.
Our approach consists of three main steps. The first
one is multi-level segmentation, which decomposes the
image to multiple segmentations from a fine level to
a coarse one. Second, we conduct a region saliency
computation step with a random forest regressor that
maps the regional features to a saliency score. Last, a
saliency map is computed by fusing the saliency maps
across multiple levels of segmentations.
The key contributions lie in the second step, region
saliency computation. Unlike most existing algorithms
that compute saliency maps heuristically from various
features and combine them to get the saliency map,
which we call saliency integration, we learn a random
forest regressor that directly maps the feature vector
of each region to a saliency score, which we call dis-
criminative regional feature integration (DRFI). This
is a principle way in image classification [19], but rarely
studied in salient object detection. It turns out that the
learnt regressor is able to automatically pick discrimi-
native features rather than heuristically hand-crafting
special features for saliency. On the other hand, we
also introduce a new descriptor, called backgroundness,
to discriminate the background from the object, which
can be considered as a counterpart of the objectness
descriptor [2].
1.1. Related work
The following gives a review of salient object detec-
tion (segmentation) algorithms that are related to our
approach. A comprehensive survey of salient object
detection can be found from [9]. The review on visual
attention modeling [7] also includes some analysis on
salient object detection.
2013 IEEE Conference on Computer Vision and Pattern Recognition
1063-6919/13 $26.00 © 2013 IEEE
DOI 10.1109/CVPR.2013.271
2081
2013 IEEE Conference on Computer Vision and Pattern Recognition
1063-6919/13 $26.00 © 2013 IEEE
DOI 10.1109/CVPR.2013.271
2081
2013 IEEE Conference on Computer Vision and Pattern Recognition
1063-6919/13 $26.00 © 2013 IEEE
DOI 10.1109/CVPR.2013.271
2083
The basis of most saliency detection algorithms can
date back to the feature integration theory [43] which
posits that different kinds of attention are responsi-
ble for binding various features into consciously ex-
perienced wholes. Later, a computational attention
model built on a biologically-plausible architecture is
proposed in [28] and completely implemented in [22].
It represents the input image from the color, intensity
and orientation channels, and computes three conspicu-
ity (saliency) maps using center-surround differences,
which are combined together to form the final master
saliency map.
Recently, a lot of research efforts have been made to
design various saliency features characterizing salient
objects or regions. Most works essentially follow
the center-surround difference (or contrast) framework.
The discriminant center-surround hypothesis is ana-
lyzed in [15, 16]. Color histograms, computed to repre-
sent the center and the surround, are used to evaluate
the center-surround dissimilarity [31]. An information
theory perspective is introduced to yield a sound math-
ematical formulation, computing the center-surround
divergence based on feature statistics [27].
The center-surround difference framework is also in-
vestigated to compute the saliency from region-based
image representation. In [23], the difference between
the color histogram of a region and its immediately
neighboring regions are used to evaluate the saliency
score. The global contrast based approach [11], com-
puting the saliency map by comparing each region with
others, aims to directly compute the global uniqueness.
Based on the regional contrast, element color unique-
ness and spatial distribution are introduced to evaluate
the saliency scores of regions [38]. The saliency map is
generated by propagating the saliency scores of regions
to the pixels.
Many other models are also proposed for saliency
computation. Center-bias, i.e. the salient object usu-
ally lies in the center of an image, is investigated
in [23, 50]. Object prior, such as connectivity prior [45],
concavity context [34], auto-context cue [48], and the
background prior [53] are also studied for saliency com-
putation. Example-based approaches, searching for
similar images of the input, are developed for salient
object detection [35, 49]. A graphical model is pro-
posed to fuse generic objectness and visual saliency to-
gether to detect objects [10]. A low rank matrix recov-
ery scheme is proposed for salient object detection [41].
A top-down approach via joint conditional random
fields and dictionary learning is introduced [54]. The
stereopsis is leveraged for saliency analysis [37]. Be-
sides, spectral analysis in the frequency domain is used
to detect salient regions [1, 20]
Additionally, there are several works directly check-
ing if an image window contains an object. The generic
objectness measure is defined by combining several im-
age cues to quantify the possibility that a window con-
tains an object [2]. A category independent object de-
tection cascade, which uses superpixel boundary inte-
gral, edge distribution and window symmetry to de-
scribe objectness, is learnt to rank a number of object
window candidates [39]. Salient object detection by
composition [13] checks if the content within an win-
dow can be composed by neighbor regions. A random
forest regression approach is adopted to directly regress
the object rectangle from the saliency map [50].
Eye fixation prediction, another visual saliency re-
search direction, also attracts a lot of interests [7, 24].
Recent developments include using isocentric curved-
ness and color [44], adopting image histogram [32],
quaternion-based spectral analysis [40], utilizing depth
cues [30], multitask sparsity pursuit [29], statistically
modeling [42], exploring patch rarities [6], combing
bottom-up and top-down features [5], task-specific vi-
sual attention [8] and so on. There are some other
saliency definitions, e.g. context-aware saliency detec-
tion [18] aiming to detect the image regions that rep-
resent the scene.
Our proposed approach differs from existing algo-
rithms on two points. In term of the saliency features,
we compute a contrast vector instead of a contrast
value used in the existing algorithms for a region. Par-
ticularly, a novel feature vector is introduced to charac-
terize the background. Our approach is also unique in
the learning strategy. In contrast to existing learning
algorithms that perform saliency integration by com-
bining saliency maps computed from different types of
features, e.g. [2, 10, 31], our approach learns to directly
integrate feature vectors to compute the saliency map.
The closely related approach [26] which also learns to
integrate the saliency features is a pixel-based algo-
rithm, while our approach is region-based that per-
forms multi-level estimation and can capture non-local
contrast. Moreover, we introduce a novel regional fea-
ture vector to characterize the background. Another
one [36] touches the discriminative feature integration
lightly without presenting a deep investigation and it
only considers the regional property descriptor. The
recent learning approach [33] aims to predict eye fixa-
tion, while our approach is for salient object detection
and moreover, we solve the problem by introducing and
exploring multi-level regional descriptors. The discrim-
inative feature fusion has also been studied in image
classification [14], which learns the adaptive weights
of features according to the classification task to bet-
ter distinguish one class from others. Our approach
208220822084
剩余7页未读,继续阅读
资源评论
keltloja
- 粉丝: 0
- 资源: 5
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- 构建数据库故障监控策略:技术实现与代码实践
- 【windows实用工具】+【统计文件夹内所有文件(包含文件夹)大小】+【方便我们管理文件 】
- AI+Go 打造你的智能办公助手
- 使用SDSOC和Xilinx Ultrascale+ZCU102平台在嵌入式操作系统中部署 CNN 加速器+项目源码+文档说明
- 个人汽车销售工作计划.docx
- 服装销售店长工作计划范文.docx
- 70154兔兔答题源码一款前后端开源的移动端答题系统
- 电视节目项目策划书.docx
- 微商防伪防窜货溯源查询系统源码二维码查询导入导出/扫码发货/物流查询
- 《测绘程序设计》第49章 利用五点光滑法进行曲线拟合案例,C# winform
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
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功