Salient Region Detection and Segmentation
Radhakrishna Achanta,Francisco Estrada,Patricia Wils,and Sabine S¨u sstrunk School of Computer and Communication Sciences(I&C),
Ecole Polytechnique F´e d´e rale de Lausanne(EPFL), {radhakrishna.achanta,francisco.estrada,patricia.wils,sabine.susstrunk}@
epfl.ch
ivrg.epfl.ch/
Abstract.Detection of salient image regions is useful for applications
like image segmentation,adaptive compression,and region-based image
retrieval.In this paper we present a novel method to determine salient
regions in images using low-level features of luminance and color.The
method is fast,easy to implement and generates high quality saliency
maps of the same size and resolution as the input image.We demonstrate
the use of the algorithm in the segmentation of semantically meaningful
whole objects from digital images.
Key words:Salient regions,low-level features,segmentation
1Introduction
Identifying visually salient regions is useful in applications such as object based image retrieval,adaptive content delivery[11,12],adaptive region-of-interest based image compression,and smart image resizing[2].We identify salient re-gions as those regions of an image that are visually more conspicuous by virtue of their contrast with respect to surrounding regions.Similar definitions of saliency exist in literature where saliency in images is referred to as local contrast[9,11].
Our method forfinding salient regions uses a contrast determinationfilter that operates at various scales to generate saliency maps containing“saliency values”per pixel.Combined,these individual maps result in ourfinal saliency map.We demonstrate the use of thefinal saliency map in segmenting whole objects with the aid of a relatively simple segmentation technique.The novelty of our approach
lies infinding high quality saliency maps of the same size and resolution as the input image and their use in segmenting whole objects.The method is effective on a wide range of images including those of paintings,video frames,and images containing noise.
The paper is organized as follows.The relevant state of the art in salient region detection is presented in Section2.Our algorithm for detection of salient regions and its use in segmenting salient objects is explained in Section3.The parameters used in our algorithm,the results of saliency map generation,seg-mentation,and comparisons against the method of Itti et al.[9]are given in Section4.Finally,in Section5conclusions are presented.
实验场
2Authors Suppressed Due to Excessive Length
2Approaches for Saliency Detection
The approaches for determining low-level saliency can be based on biological models or purely computational ones.Some approaches consider saliency over several scales while others operate on a single scale.In general,all methods use some means of determining local contrast of image regions with their sur-roundings using one or more of the features of color,intensity,and orientation. Usually,separate feature maps are created for each of the features used and then combined[8,11,6,4]
to obtain thefinal saliency map.A complete survey of all saliency detection and segmentation research is beyond the scope of this paper,here we discuss those approaches in saliency detection and saliency-based segmentation that are most relevant to our work.
Ma and Zhang[11]propose a local contrast-based method for generating saliency maps that operates at a single scale and is not based on any biological model.The input to this local contrast-based map is a resized and color quan-tized CIELuv image,sub-divided into pixel blocks.The saliency map is obtained from summing up differences of image pixels with their respective surrounding pixels in a small neighborhood.This framework extracts the points and regions of attention.A fuzzy-growing method then segments salient regions from the saliency map.
Hu et al.[6]create saliency maps by thresholding the color,intensity,and orientation maps using histogram entropy thresholding analysis instead of a scale space approach.They then use a spatial compactness measure,computed as the area of the convex hull encompassing the salient region,and saliency density, which is a function of the magnitudes of saliency values in the saliency feature maps,to weigh the individual saliency maps before combining them.
Itti et al.[9]have built a computational model of saliency-based spatial at-tention derived from a biolog
ically plausible architecture.They compute saliency maps for features of luminance,color,and orientation at different scales that ag-gregate and combine information about each location in an image and feed into a combined saliency map in a bottom-up manner.The saliency maps produced by Itti’s approach have been used by other researchers for applications like adapting images on small devices[3]and unsupervised object segmentation[5,10].
Segmentation using Itti’s saliency maps(a480x320pixel image generates a saliency map of size30x20pixels)or any other sub-sampled saliency map from a different method requires complex approaches.For instance,a Markov random field model is used to integrate the seed values from the saliency map along with low-level features of color,texture,and edges to grow the salient object regions [5].Ko and Nam[10],on the other hand,use a Support Vector Machine trained on the features of image segments to select the salient regions of interest from the image,which are then clustered to extract the salient objects.We show that using our saliency maps,salient object segmentation is possible without needing such complex segmentation algorithms.
Recently,Frintrop et al.[4]used integral images[14]in VOCUS(Visual Object Detection with a Computational Attention System)to speed up com-putation of center-surround differences forfinding salient regions using separate
Salient Region Detection and Segmentation3 feature maps of color,intensity,and orientation.Although they obtain better resolution saliency maps as compared to Itti’s method,they resize the feature saliency maps to a lower scale,thereby losing resolution.We use integral images in our approach but we resize thefilter at each scale instead of the image and thus maintain the same resolution as the original image at all scales.
3Salient region detection and segmentation
This section presents details of our approach for saliency determination and its use in segmenting whole objects.An overview of the complete algorithm is presented in Figure1.Using the saliency calculation method described later, saliency maps are created at different scales.These maps are added pixel-wise to get thefinal saliency maps.The input image is then over-segmented and the segments whose average saliency exceeds a certain threshold are chosen.轻轻的一声叮咛
Fig.1.Overview of the process offinding salient regions.(a)Input image.(b)Saliency maps at different scales are computed,added pixel-wise,and normalized to get thefinal saliency map.(c)Thefinal saliency map and the segmented image.(d)The output image containing the salient object that is made of only those segments that have an average saliency value greater than the threshold T(given in Section3.1).
3.1Saliency calculation
In our work,saliency is determined as the local contrast of an image region with respect to its neighborhood at various scales.This is evaluated as the distance between the average feature vector of the pixels of an image sub-region with the average feature vector of the pixels of its neighborhood.This allows obtaining a combined feature map at a given scale by using feature vectors for each pixel, instead of combining separate saliency maps for scalar values of each feature.At a given scale,the contrast based saliency value c i,j for a pixel at position(i,j)in the image is determined as the distance D between the average vectors of pixel
4Authors Suppressed Due to Excessive
Length
Fig.2.(a)Contrast detectionfilter showing inner square region R1and outer square region R2.(b)The width of R1remains constant while that of R2ranges according to Equation3by halving it for each new scale.(c)Filtering the image at one of the scales in a raster scan fashion.
features of the inner region R1and that of the outer region R2(Figure2)as:
c i,j=D
1
N1
N1
p=1
v p
火成岩,
1
N2
N2
q=1
v q
(1)
where N1and N2are the number of pixels in R1and R2respectively,and v is the vector of feature elements corresponding to a pixel.The distance D is a Euclidean distance if v is a vector of uncorrelated feature elements,and it is a Mahalanobis distance(or any other suitable distance measure)if the elements of the vector are correlated.In this work,we use the CIELab color space[7],a
ssuming sRGB images,to generate feature vectors for color and luminance.Since perceptual differences in CIELab color space are approximately Euclidian,D in Equation 1is:
c i,j= v1−v2 (2) where v1=[L1,a1,b1]T an
d v2=[L2,a2,b2]T ar
e the average vectors for regions R1and R2,respectively.Since only average feature vector values o
f R1and R2need to be found,we use the integral image approach as used in[14]for computational efficiency.A change in scale is affected by scalin
g the region R2 instead of scaling the image.Scaling thefilter instead of the image allows the generation of saliency maps of the same size and resolution as the input image. Region R1is usually chosen to be one pixel.If the image is noisy(for instance if hig
h ISO values are used when capturing images,as can often be determined with the help of Exif data(Exchangeable File Information Format[1])then R1 can be a small region of N×N pixels(in Figure5(f)N is9).
For an image of width w pixels and height h pixels,the width of region R2,
namely w R
2is varied as:
w
2
≥(w R
2
)≥
w
8
(3)
assuming w to be smaller than h(else we choose h to decide the dimensions of R2).This is based on
the observation that the largest size of R2and the smaller ones(smaller than w/8)are of less use infinding salient regions(see Figure 3).The former might highlight non-salient regions as salient,while the latter are basically edge detectors.So for each image,filtering is performed at three
钼制品
Salient Region Detection and Segmentation5
Fig.3.From left to right,original image followed byfiltered images.Filtering is done using R1of size one pixel and varying width of R2.When R2has the maximum width,certain non salient parts are also highlighted(the ground for instance).It is the saliency maps at the intermediate scales that consistently highlight salient regions. The last three images on the right mainly show edges.济南丝足
different scales(according to Eq.3)and thefinal saliency map is determined as a sum of saliency values across the scales S:
c i,j(4)
m i,j=
S
∀i∈[1,w],j∈[1,h]where m i,j is an element of the combined saliency map M obtained by point-wise summation of saliency values across the scales.
3.2Whole Object Segmentation using Saliency Maps
The image is over-segmented using a simple K-means algorithm.The K seeds for the K-means segmentation are automatically determined using the hill-climbing algorithm[13]in the three-dimensional CIELab histogram of the image.The
安乐死论文
Fig.4.(a)Finding peaks in a histogram using a search window like(b)for a one dimensional histogram.
hill-climbing algorithm can be seen as a search window being run across the space of the d-dimensi
onal histogram tofind the largest bin within that window. Figure4explains the algorithm for a one-dimensional case.Since the CIELab feature space is three-dimensional,each bin in the color histogram has3d−1=26 neighbors where d is the number of dimensions of the feature space.The number
豫公网安备41160202000603
站长QQ:729038198
关于我们
投诉建议