在L^*a^*b^*颜色空间中通过相干聚类进行图像分割资源-CSDN文库

49 浏览量 2021-02-23 13:52:51 上传评论收藏 3.52MB PDF 举报

在图像处理和计算机视觉领域中，图像分割是一项基础而重要的任务，其目的是将图像划分为多个区域或对象，便于后续分析、识别和处理。图像分割方法多种多样，其中基于聚类的方法因其简洁性和直观性而被广泛研究。本文重点介绍了在L*a*b*颜色空间中通过相干聚类算法进行图像分割的研究。要理解L*a*b*颜色空间。这是一种被广泛认可的色彩模型，用于描述色彩。在这个模型中，L*代表亮度（Luminance），a*和b*代表色彩轴，其中a*为从绿到红的色度，b*为从蓝到黄的色度。这种颜色空间之所以适合图像分割，是因为它更接近人眼对颜色的感知方式，并且能够减少颜色通道间的依赖关系，使得分割算法更加有效。接着，我们来讨论相干聚类算法。相干聚类算法的主要目的是在特征空间中保持良好的数据一致性，这对于聚类的鲁棒性至关重要。文章提到，传统的聚类算法在特征空间中分布复杂时，往往难以得到鲁棒的聚类结果。因此，文章提出了一种新的鲁棒性聚类算法，并将其应用于L*a*b*颜色空间中的像素特征聚类。通过这种方法，每个像素都可以直接设置为其对应聚类的成员，从而简单地完成图像分割。文章还提到了基于最小描述长度（Minimum Description Length, MDL）理论的参数自适应选择方法。最小描述长度原理是一种用于模型选择的方法，它基于奥卡姆剃刀原则，倾向于选择最简洁的模型来解释数据。在这个场景下，它帮助自动选择图像分割方法的最佳参数，使得分割结果既准确又高效。此外，文章还提到半监督判别分析在图像分割中的应用。半监督学习是一种结合了有标签数据和无标签数据的学习方式，在仅有少量有标签数据的情况下尤其有用。半监督判别分析能够更好地利用未标记数据，提高聚类算法在特征空间中的鲁棒性。文章所提出的图像分割方法在伯克利分割数据库（Berkeley Segmentation Database）上进行了测试。伯克利分割数据库是一个广泛使用的公共图像分割基准测试平台，它包含大量图像及其手工分割结果。通过在这个数据库上的测试，文章的方法不仅证明了其有效性，还展示了其在实际应用中分割质量的优越性。为了更好地理解文章内容，我们需要了解几个关键技术点： 1. 聚类算法：聚类是将数据点自动分组成多个类或群集的过程。在图像分割中，聚类算法能够根据像素特征将图像中的不同区域分隔开来。 2. 半监督学习：在机器学习中，半监督学习是介于监督学习和无监督学习之间的一种方法。它使用少量的标记数据和大量的未标记数据来进行训练和预测，旨在提高学习效率和结果的准确性。 3. 最小描述长度（MDL）：它是一种模型选择准则，用于评价一个模型的复杂度与该模型对数据的拟合程度之间的权衡。在图像分割中应用MDL理论可以帮助选择最优的分割参数。总结来说，本研究通过提出一种适用于L*a*b*颜色空间的鲁棒性相干聚类算法，并结合半监督判别分析和MDL原理，实现了一种高效的图像自动分割方法。该方法在实际的图像分割测试中表现优异，为计算机视觉和图像处理领域提供了有价值的参考。

资源推荐

资源详情

资源评论

Image segmentation via coherent clustering in L

⁄

color space

Rui Huang

⇑

, Nong Sang

, Dapeng Luo

, Qiling Tang

China Ship Development and Design Center, 268, Zi-Yang Road, Wuhan 430064, PR China

Institute for Pattern Recognition and Artiﬁcial Intelligence, Huazhong University of Science and Technology, 1037, Luo-Yu Road, Wuhan 430074, PR China

College of Mechanical and Electronic Information, China University of Geosciences, 388, Lumo Road, Wuhan 430074, PR China

College of Biomedical Engineering, South-Central University For Nationalities, 708, Min-Yuan Road, Wuhan 430074, PR China

article info

Article history:

Received 21 February 2010

Available online 1 February 2011

Communicated by Y.J. Zhang

Keywords:

Image segmentation

Superpixel

Semi-supervised discriminant analysis

Clustering

abstract

Automatic image segmentation is always a fundamental but challenging problem in computer vision. The

simplest approach to image segmentation may be clustering feature vectors of pixels at ﬁrst, then label-

ing each pixel with its corresponding cluster. This requires that the clustering on feature space must be

robust. However, most of popular clustering algorithms could not obtain a robust clustering result yet, if

the clusters in feature space have a complex distribution. Generally, for most of clustering-based segmen-

tation methods, it still needs more constraints of positional relations between pixels in image lattice to be

utilized during the procedure of clustering. Our works in this paper address the problem of image seg-

mentation under the paradigm of pure clustering-then-labeling. A robust clustering algorithm which could

maintain good coherence of data in feature space is proposed and utilized to do clustering on the L

⁄

color feature space of pixels. Image segmentation is straightforwardly obtained by setting each pixel with

its corresponding cluster. Further, based on the theory of Minimum Description Length, an effective

approach to automatic parameter selection for our segmentation method is also proposed. We test our

segmentation method on Berkeley segmentation database, and the experimental results show that our

method compares favorably against some state-of-the-art segmentation methods.

1. Introduction

Automatic image segmentation, aiming at partitioning an image

into homogeneous and meaningful regions, is a fundamental

problem in the ﬁelds of image processing and computer vision. Seg-

mentation provides a compact region-based description for an im-

age, which could be adopted in designing more complex algorithms

in many high-level vision applications, such as recognition, image

retrieval and scene classiﬁcation.

Casting image segmentation as a clustering problem might be the

most popular and straightforward approach. Ideally, a feature vector

is extracted at each pixel via computing its local properties, such as

color and/or texture. The feature space is then clustered, and each

pixel is labeled with the cluster label corresponding to its feature.

However, data in feature space do not always maintain the compact

coherence (i.e. the distributions of clusters in feature space do not al-

ways have clear cut-off boundaries between each other). Simply

using the general clustering methods (i.e. K-means), would not ob-

tain a coherent or robust clustering result. To get satisfying segmen-

tation results, more ad hoc post-processing may be needed.

Instead of purely clustering on color and/or texture feature

space, many successful clustering-based segmentation methods

have incorporated the positional information of pixels in image lat-

tice into the procedure of clustering. In mean-shift based image

segmentation system (Comaniciu and Meer, 2002), during the clus-

tering, besides the feature bandwidth parameter h

, authors also

introduced another spatial constraint parameter h

, known as the

spatial bandwidth. In normalize-cut image segmentation system

(Shi and Malik, 2000), when computing the pairwise weight be-

tween pixels, both information of pixels in feature space and in im-

age lattice were combined. Recently, Yang et al. proposed a new

segmentation method (Yang et al., 2008) based on the Lossy Data

Coding and Compression theory (Ma et al., 2007). To enforce the seg-

mentation results having good spatial continuity among segments,

they maintained a Region Adjacency Graph (RAG) during the

agglomerative clustering process.

Among the researches on image segmentation in the past years,

there are also many other techniques. These include region split-

and-merge methods (Deng et al., 1999; Zhu and Yuille, 1996), Mar-

kov random ﬁeld (MRF) based approaches (Panjwani and Healey,

1995; Tu and Zhu, 2002), mixture model based methods (Sanjay-

Gopal and Hebert, 1998; Nikou et al., 2007), non-parametric prob-

abilistic method (Andreetto et al., 2007), and graph theoretic ap-

proaches (Felzenszwalb and Huttenlocher, 2004; Wang et al.,

2008), etc. Although these techniques have achieved signiﬁcant

improvement, and enlightened us on image segmentation with

many useful ideas, they require us to address many sophisticated

issues, such as more complex feature extraction, more complex

statistical modeling, and more complex optimization techniques.

doi:10.1016/j.patrec.2011.01.013

⇑

Corresponding author. Tel.: +86 27 88076892; fax: +86 27 88042076.

E-mail address: ruihuang2008@gmail.com (R. Huang).

Pattern Recognition Letters 32 (2011) 891–902

Contents lists available at ScienceDirect

Pattern Recognition Letters

journal homepage: www.elsevier.com/locate/patrec

Our work in this paper, withdraws image segmentation back to

the paradigm of pure clustering-then-labeling. Specially, instead of

processing on pixel level, we complete feature extraction and

clustering on the superpixel (Ren and Malik, 2003; Mori, 2005) le-

vel. Image is ﬁrst over-segmented into many small homogeneous

regions, known as superpixels. On feature extraction of each

superpixel, we adopt a simple 3-D color feature vector which is

simply computed by averaging the L

⁄

color values over all of

the pixels in each superpixel. On clustering all superpixels in this

feature space, we propose a novel unsupervised clustering algo-

rithm. We combine mean-shift (Comaniciu and Meer, 2002) clus-

tering algorithm with Semi-supervised Discriminant Analysis

(SDA) (Cai et al., 2007) in an incremental learning scheme. We

use mean-shift to generate the cluster label, and use SDA to do

subspace selection. Both these steps are performed alternately.

Experimental results show that clustering result obtained by our

clustering algorithm is more coherent than those obtained by sev-

eral other popular clustering algorithms. Due to the robustness of

our clustering algorithm, the segmentation of image could ob-

tained by directly setting the superpixels on image with their cor-

responding clustering labels. Compared with several previous

methods, our segmentation method not only is simple and

straightforward, also has a competitive performance in quantita-

tive evaluation on Berkeley segmentation database (Martin

et al., 2001).

Another important work in this paper, is our proposed approach

to automatic parameter selection for our segmentation method.

Usually, given a segmentation algorithm, it’s not hard to select a

proper parameter of this algorithm via hand-tuning to obtain a

good segmentation for an image. But if we want to segment well

a database of images, say the Berkeley segmentation database,

and make the benchmark comparison with several segmentation

methods, we must build a methodological approach to parameter

selection. Under the same conditions, the aim is that the approach

can automatically select the proper parameter of segmentation

method for every image in the database to obtain a good segmenta-

tion. In this paper, from the experiments on the 300 images in

Berkeley segmentation database, it showed that our parameter

selection method had well achieved this aim.

2. Color feature of superpixel

Conventionally, image segmentation is thought as a low-level

vision task in which feature is usually extracted for each pixel.

For an image with a typical size, hundreds of thousands of data

points will have to be processed if we directly do clustering in

the feature space. In this paper, to reduce the computational cost

of clustering, we utilize the feature extracted on the superpixel le-

vel, instead of the pixel level.

Superpixel (Ren and Malik, 2003) is a small atomic region ob-

tained by doing over-segmentation on image based on the local

cues such as color (or texture) and edges. Compared with the large

number of pixels in an image, the over-segmented result usually

has only several hundreds of superpixels. Furthermore, two other

desirable properties of superpixel also motivate us to utilize it

for image segmentation. One is that each superpixel is a perceptu-

ally consistent unit, i.e. all pixels in a superpixel are almost uniform

in color or texture. The other is that the superpixel map obtained

by over-segmentation could well preserve most of the structures

(i.e. edges or contours) necessary for segmentation at the scale of

interest. In (Ren and Malik, 2003), authors have validated this

property by comparing the reconstructed segmentation from

superpixel map with the groundtruth using a contour-base mea-

sure. Both of these two properties make us believe that if we could

obtain a good clustering result in the feature space of superpixels,

and then label each superpixel with its corresponding cluster, we

can obtain a satisfying image segmentation result.

Since L

⁄

color metric was specially designed to best

approximate to human vision, and by which using the simple

Euclidean distance could differentiate among colors perceptually.

In this paper, on feature extraction of each superpixel, we ﬁrst

convert the original RGB color of image into the L

⁄

color

space, and then compute the average value of the L

⁄

color

over the pixels in each superpixel. The obtained average value

is a 3-D vector in L

⁄

color space and will work as the feature

utilized to cluster the superpixels. In later part of this paper, we

call this 3-D L

⁄

color feature as superpixel color for

convenience.

Fig. 1(a) is a natural image with a size of 180  120 from Cor-

el image set. Fig. 1(b) is the superpixel map of this image, which

is generated by using the publicly available superpixel code

(Mori, 2005). In this superpixel map, there are 1071 superpixels

in all. Each superpixel is almost a perceptually homogeneous re-

gion. Fig. 1(c) is the human labeled segmentation for Fig. 1(a).

This ground truth

contains three class of labels for each pixel

in image Fig. 1(a). Similar to the computation of our superpixel

color, we can also approximately obtain the label of each super-

pixel in Fig. 1(b) via computing the mode of labels in Fig. 1(c) over

the pixels in each superpixel. When each superpixel is equipped

with a ground truth label, we can not only clearly observe the dis-

tribution of all superpixels in our superpixel color feature space,

but also have an immediate visual evaluation to the clustering re-

sults obtained by using other clustering algorithms in this super-

pixel color space.

Fig. 1(d) illustrates the distribution of 1071 superpixels of

Fig. 1(b) in our superpixel color feature space. Each data point is

marked according to the superpixel label computed from the

ground truth in Fig. 1(c). Fig. 1(d) can be thought as a ground truth

for evaluating a clustering result on these 1071 superpixels. If we

want to obtain a better ﬁnal segmentation result for Fig. 1(a) sim-

ply by clustering-then-labeling, the clustering result in our super-

pixel color feature space should well approximate Fig. 1(d).

Fig. 1(e)–(g), respectively illustrate the clustering results on

these 1071 superpixels, which are obtained by using three popular

clustering algorithms. The clustering result shown in Fig. 1(e) is

generated by using the K-means clustering algorithm. To compare

with the ground truth in Fig. 1(d), we set K = 3. The clustering re-

sult shown in Fig. 1(f) is generated by using the normalized-cut

clustering algorithm.

This algorithm has only one tuning parame-

ter, which is the number of clusters to be generated. Here, we also

set it to be 3. The clustering result shown in Fig. 1(g) is generated

by using the mean-shift (Comaniciu and Meer, 2002) clustering

algorithm. The tuning parameter in this algorithm is called feature

bandwidth and usually denoted by h

. We set h

= 17 to make the

clustering result having three clusters.

From these clustering results in Fig. 1(e)–(g), we can see that

none of them could well approximate the ground truth in

Fig. 1(d). If we directly label the superpixel with these clustering

results on image, we will not obtain a good segmentation result

comparing with the ground truth in Fig. 1(c). This could partially

explain why the image segmentation methods in (Comaniciu and

Meer, 2002; Shi and Malik, 2000) incorporated the positional infor-

mation of pixels into their clustering process. In following sections,

we will show our clustering results on these 1071 superpixels for

comparison.

Both Fig. 1(a) and (c) can be downloaded from: www.cs.toronto.edu/~hexm/data/

corel_subset.mat.

Code can be download from www.cis.upenn.edu/~jshi/software/ﬁles/

NcutClustering_7.zip.

892 R. Huang et al. / Pattern Recognition Letters 32 (2011) 891–902

剩余11页未读，继续阅读

评论收藏

内容反馈

weixin_38685521

粉丝: 4
资源: 935

在L ^ * a ^ * b ^ *颜色空间中通过相干聚类进行图像分割

最新资源

在L ^ * a ^ * b ^ *颜色空间中通过相干聚类进行图像分割

空间一致性约束谱聚类算法用于图像分割

Opencv Kmeans聚类进行图像颜色聚类分割（绝对有用）

利用聚类方法进行图像分割

利用聚类技术实现纹理图像分割

lab空间颜色分割

基于彩色空间的二维OSTU 的图像分割算法

ColorClusterPublic:一种基于色彩聚类的图像分割方法

结合非局部信息与模糊聚类的SAR图像分割方法.docx

计算机研究 -基于近邻传播聚类的极化SAR图像分类.pdf

多特征-谱聚类的SAR图像溢油分割 (2010年)

SAR图像处理

基于层次聚类的多颜色空间融合分割方法

基于K-means聚类的图像分割

基于FCM算法的SAR图像相干斑噪声滤波算法研究 图像滤波算法.pdf

谱聚类源码matlab-crisp-boundaries:使用逐点互信息进行边界检测和图像分割

一种基于SVR的SAR图像自适应滤波算法 图像滤波算法.pdf

用于图像处理的非线性松弛标记：使用概率松弛提高 2D 单色/多光谱图像的空间相干性-matlab开发

Anomaly Detection in Medical Imaging -- A Mini Review.pdf

一种杂波分类辅助的近海岸模糊杂波抑制方法.docx

基于CORDIC的反正弦和反余弦计算的FPGA实现

BA无标度网络中的SIR模型

使用3DCNN和卷积LSTM进行手势识别学习时空特征

基于三次贝塞尔曲线的类汽车曲率连续路径平滑

基于机器学习的设备剩余寿命预测方法综述

基于维纳过程的退化模型，具有递归过滤算法，可用于估计剩余使用寿命

最新资源

基于FCM算法的SAR图像相干斑噪声滤波算法研究图像滤波算法.pdf

一种基于SVR的SAR图像自适应滤波算法图像滤波算法.pdf