844 IEEE TRANSACTIONS ON BROADCASTING, VOL. 62, NO. 4, DECEMBER 2016
for the first time and proposed a feed-forward model to com-
bine these features. The Koch and Ullman [17] model was
firstly completely implemented and verified by Itti et al. [9]
in a biologically plausible way. Since then, many models
with different assumptions for attention modeling have been
proposed. It is very difficult to classify all the saliency mod-
els. In this section, we provide a brief overview of the recent
saliency models. The models are classified into four classes
in an intuitive way: local model, global model, local+global
combined model, and others. Note that we focus on models for
human fixation prediction instead of those models that detect
the most salient region or object in an image.
Itti et al. [9] computed saliency maps for each of the
three features (e.g., colors, intensity, orientations) in paral-
lel, where each feature was computed by a set of linear
“center-surround” operations akin to visual receptive fields.
Then these maps were linearly summed and normalized to
yield the “conspicuity maps”. Le Meur et al. [18] proposed
an approach based on the understanding of the HVS behav-
ior. Contrast sensitivity functions, perceptual decomposition,
visual masking, and center-surround interactions were some
of the features implemented in this model. Itti and Baldi [19]
defined surprising stimuli as those which significantly change
beliefs of an observer by measuring the Kullback-Leibler (KL)
distance between posterior and prior beliefs of the observer.
Harel et al. [20] first extracted features similar to [9] to obtain
three multi-scale feature maps (e.g., colors, intensity, orien-
tations). Then, a fully connected graph was built and each
graph was treated as a Markov chain to build an activation
map. Seo and Milanfar [21] first computed local features mea-
suring the likeness of a pixel to its surroundings. Then, the
matrix cosine similarity (a generalization of cosine similar-
ity) was employed to measure the similarity of each pixel to
its surroundings. Gao et al. [22] defined saliency as classi-
fication with minimal expected error. The KL distance was
utilized to measure mutual information between features at
a scene location and class labels. Higher mutual information
between a region and class of interest indicates higher saliency
of that region. Gu et al. [23] introduced free energy theory into
saliency detection. They computed the local entropy of the gap
between an image and its predicted version reconstructed from
the input one by a semi-parametric model, which fused the
parametric autoregressive (AR) operator that can simulate a
broad range of natural scenes and the non-parametric bilateral
filtering that works stably at image edges.
Bruce and Tsotsos [7] proposed the Attention based on
Information Maximization (AIM) model aiming at maximiz-
ing information sampled from a scene. The proposed operation
was based on Shannon’s self-information measure and was
implemented in a neural circuit, which was demonstrated to
have close ties with the circuitry existent in the primate visual
cortex. Zhang et al. [24] proposed a definition of saliency
by considering what the visual system was trying to opti-
mize when directing attention. The resulting model was a
Bayesian framework from which bottom-up saliency emerged
naturally as the self-information of visual features, and overall
saliency (incorporating top-down information with bottom-
up saliency) emerged as the pointwise mutual information
between the features and the target when searching for a
target. Garcia-Diaz et al. [25] proposed a visual saliency
approach that relied on a contextually adapted representation
produced through adaptive whitening of color and scale fea-
tures. The proposal was grounded on the specific adaptation
of the basis of low-level features to the statistical structure
of the image. Adaptation was achieved through decorrelation
and contrast normalization in several steps in a hierarchical
approach, in compliance with coarse features described in bio-
logical visual systems. Saliency was simply the square of the
vector norm in the resulting representation. Riche et al. [13]
proposed a saliency prediction model, which selected infor-
mation worthy of attention based on multi-scale spatial rar-
ity. First, they extracted low-level color and medium-level
orientation features. Afterwards, a multi-scale rarity mecha-
nism was applied. Finally, they fused rarity maps into final
saliency map.
Li et al. [26] proposed a saliency model by combining
global information from frequency domain analysis and local
information from spatial domain analysis. In frequency domain
analysis, they suppressed repeating patterns by using spectrum
smoothing. In spatial domain analysis, they enhanced regions
by using a center-surround mechanism. Goferman et al. [15]
proposed a context-aware saliency detection model based on
four principles. First, local low-level considerations such as
color and contrast. Second, global considerations which sup-
press frequently occurring features while maintaining features
that deviate from the norm. Third, visual organization rules
which state that visual forms may possess one or several
centers of gravity about which the form is organized. Four,
high-level factors, such as human faces. Borji and Itti [14]
proposed a framework that measured patch rarities locally and
globally in RGB and Lab color space and fused local and
global saliency maps of all channels from both color spaces
into the final saliency map. Liu et al. [27] measured visual
saliency as the unpredicted information of image patch through
an order-adaptive predictor under minimum description length
principle. Furthermore, a structural redundancy operator was
also involved to improve the saliency detection performance.
Hou and Zhang [10] proposed the spectral residual saliency
model based on the idea that statistical singularities in the
spectrum might be responsible for anomalous regions in the
image, where proto objects were popped up. They first ana-
lyzed the log spectrum of each image and obtained the spectral
residual. Then they transformed spectral residual to spatial
domain to obtain the saliency map. Guo et al. [28]showed
that it was the phase spectrum, not the amplitude spectrum, of
the Fourier transform that was the key in obtaining the loca-
tion of salient areas. Later, Guo and Zhang [29] proposed a
quaternion representation of an image which was composed of
intensity, color, and motion features. Based on the principle of
the phase spectrum of Fourier transform, the spatiotemporal
saliency map was calculated by its quaternion representation.
Achanta et al. [30] proposed a frequency-tuned approach using
low-level features of color and luminance. The input RGB
image was transformed to Lab color space and blurred with
a Gaussian kernel to eliminate noise and texture details. Then
the saliency map was computed using the Euclidean distance
剩余12页未读，继续阅读
weixin_38737144
- 粉丝: 4
- 资源: 942
我的内容管理
展开
最新资源
- zlib-1.2.12压缩包解析与技术要点
- 微信小程序滑动选项卡源码模版发布
- Unity虚拟人物唇同步插件Oculus Lipsync介绍
- Nginx 1.18.0版本WinSW自动安装与管理指南
- Java Swing和JDBC实现的ATM系统源码解析
- 掌握Spark Streaming与Maven集成的分布式大数据处理
- 深入学习推荐系统：教程、案例与项目实践
- Web开发者必备的取色工具软件介绍
- C语言实现李春葆数据结构实验程序
- 超市管理系统开发：asp+SQL Server 2005实战
- Redis伪集群搭建教程与实践
- 掌握网络活动细节：Wireshark v3.6.3网络嗅探工具详解
- 全面掌握美赛：建模、分析与编程实现教程
- Java图书馆系统完整项目源码及SQL文件解析
- PCtoLCD2002软件：高效图片和字符取模转换
- Java开发的体育赛事在线购票系统源码分析
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
