面向区域图像压缩的区域分割方法资源-CSDN文库

52 浏览量 2021-03-28 00:43:47 上传评论收藏 671KB PDF 举报

资源推荐

资源详情

资源评论

A region segmentation method for region-oriented image compression

Rongchang Zhao

a,b,

, Yide Ma

Northwest University for Nationalities, College of Electrical Engineering, No. 1 Xibei Xincun, Lanzhou, Gansu, PR China

School of Information Science & Engineering, Lanzhou University, Lanzhou, Gansu, PR China

article info

Article history:

Received 15 January 2011

Received in revised form

28 January 2012

Accepted 31 January 2012

Communicated by X. Gao

Available online 22 February 2012

Keywords:

Image processing

Segmentation

Segmented image coding

Neural network

Fitting function

abstract

In order to obtain homogeneous regions and smooth contours for region-oriented image compression,

gradient-coupled spiking cortex model is designed and applied to digital image segmentation. Inspired

by the knowledge of visual cortex, the model is composed of neurons with spike coupling and gradient

enhancement, and it is same as the one in the visual cortex which can distinguish some objects in real

scene through capturing boundary information. The model smoothes pixels within regions and

enhances pixels at boundaries by creating a ﬁtting function. Outputs of the model are the desired

segmented image after connection components label. Experiments show that the method not only

detects regions of original image, but also remains succinct effective contours, so it is suitable for

region-oriented image compression.

1. Introduction

Digital image processing is a subject about 2D discrete signal

processing in which a digital image is represented as a matrix and

the value is called as gray intensity. Image segmentation is to

divide an image into some connected components based on the

location and its gray intensities. While image compression is to

represent images with the shorter bits and the more information.

Region-oriented image compression technique handles images

based on regions as contours of objects but regular rectangles as

processing units.

Segmented image coding (SIC) is considered as one of region-

oriented image compression technology [1], which divides an

input image into two parts: contours and regions with slowly

varying image intensity. The contours are coded with a method

while the homogeneous regions are represented by linear combi-

nation of orthogonal basis functions. Here, segmentation of

regions is one of key problems to be solved. The actual perfor-

mance of image compression depends highly on the segmentation

algorithm. As proposed by Christopoulos et al. [2], the segmented

image is expected as the one that has the controlled number of

regions, perfect homogeneity within a region, less small-region

and smooth contours.

In order to represent images effectively, segmented regions are

expected as homogeneous as possible, and the number of small

regions should be limited [3]. Furthermore, many experiments

show that quite a few of bits are spent for coding contours in SIC.

Hence, the number of contour pixels is essential for compression

ratio of the image compression method [4]. The segmented

algorithm should not only classify similar neighbor pixels into

same regions, but also contour objects with the least pixels. Those

properties are essential but no existing methods could meet.

Generally, there are two approaches [5] to partition an image

into regions: region-based segmentation and edge detection. For

region-based segmentation, all pixels with similar attributes are

grouped together and marked as a region. Pixels are selected

based on the similarity of some attributes, for a gray image, the

basic attributes are gray intensity and spatial distance. Pixels may

be partitioned into the same region if they have similar gray

intensity and adjacent space distance. Edges are crucial local

information of regions and they generally occur at the common

border of two or more regions. Edge detection aims to form a

boundary to separate different regions and it implements through

looking for the discontinuity of the image intensity. There are

many edge detection algorithms such as Canny, Sobel, Prewitt,

Laplace and so on [5], but edges detected by this method always

are not closed.

Splitting and merging technology [6] is a region-based method,

which starts with a unit in an image, and then the units split and/or

merge together with some criteria until the desired result. This

approach could obtain preferable segmented image but the process

of splitting or merging is difﬁcult to control. For region-oriented

Contents lists available at SciVerse ScienceDirect

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

Neurocomputing

doi:10.1016/j.neucom.2012.01.007

Corresponding author at: Northwest University for Nationalities, College

of Electrical Engineering, No. 1 Xibei Xincun, Lanzhou, Gansu, PR China.

Tel.: þ86 931 8912786.

E-mail address: Byrons.zhao@gmail.com (R.C. Zhao).

Neurocomputing 85 (2012) 45–52

image compression, Christopoulos et al. [2] proposed a segmented

method based on splitting and merging technology. The method

produces an over-segmented image ﬁrstly by splitting, and then

merges some regions based on the difference of their gray mean

intensity and a cost function. The cost function is founded on the

gradient information, the size of the segments and the shared

contour length of adjacent segments. The over-segmented image

is to reduce the risk of losing important edges and the merging is

to classify some small segments into its neighboring regions, but

the merging is unruly, and the optimal segmented image can be

obtained by trial and error. Furthermore, the segmented image

contains many unconsidered tiny texture or noise and it is a

contradiction because if the over-segmented image is excessive,

the important edge is captured but more useless textures are

contained. Contrarily, the important edge information is lost. The

clustering method [7–9] is another region-based approach and it

regards image segmentation as common data classiﬁcation to

process through feature extraction and decision [10,11], and the

decision step provides ﬁnal segmented images, therefore, the

deﬁnition of feature space is signiﬁcant and difﬁcult for different

kinds of images.

Active contour model is an effective edge detection method,

which is based on variational method and be widely applied to

SAR, medical image segmentation and so on. At present, depend-

ing on the curve expression, active contour model can be

classiﬁed as parametric active contour model and geometric

active contour model. Those models achieve its segmentation

through minimizing the objective energy function and evolving

the initial curve toward the edges of objects. The C-V model [12]

is a classical geometric active contour model based on level set

and curve evolution and it takes full advantage of global gray

intensity, but the result is not good because the edge positioning

accuracy is not high. In recent years, various models have been

proposed, for example, active contour with shape prior [13], fast

global minimization of active contour model [14], GVF snake [15]

and methods based on level set [16].

Furthermore, a variety of hybrid methods appear; for example,

watershed method [17] which considers the border of local

minima as watershed and segments the regions into catchment

basins. It is better at the weak border, so the over-segmentation

may emerge where there is noise. Graph cut [18], which considers

an image as graph with nodes and edges, received considerable

attention as a method for image segmentation. It could separate

the object from a scene but the background is not homogeneous

because some textures still hide in. Further, contours are not

smooth enough. In [19], minimum description length principle is

introduced into image segmentation. Images are modeled as

Gaussian random ﬁelds with piece-wise homogeneous mean

and variance. Based on the changes in the mean and/or variance,

the edges can be detected.

For the past few years, artiﬁcial neural network has already

became a well-known technology used in computer vision, of

course, it is widely used in image segmentation. Motivated by a

histogram clustering approach to image segmentation, Buhmann

et al. [20] proposed a network of leaky integrate-and-ﬁre neurons

to segment gray image. The ﬁring rate of class neurons is

employed to encode image segmentation because the connection

between neighboring neurons can smooth adjacent similar neigh-

bors. Meftah et al. [21] applied spiking neural network model for

image segmentation and edge detection, and addressed the issue

of parameter selection by an unsupervised learning method.

In this paper, gradient-coupled spiking cortex model is pro-

posed to segment gray images into homogeneous regions and

smooth contours for SIC. The model is composed of neurons with

spike coupling and gradient enhancement, and neurons imitate

the ones in the visual cortex which can capture boundary

information of the scene. Compared with others, the model

smoothes pixels within regions and enhances pixels at boundaries

by creating a ﬁtting function, so it could eliminate some noise or

disturbances and emphasize functions of the similar neighbors on

ﬁnding discontinuous of stimulus. Then, the stimulus is segmen-

ted by a series of dynamic thresholds and ﬁtting curved-surfaces.

The stimuli coupled with the spikes of neighbors turn into the

internal state, those make a curved-surface and the dynamic

threshold forms another curved-surface. At different times, trans-

ection of the multilayer curved-surfaces produces a series of

binary matrixes which contain the information of edge, region

and texture. The time matrix of the neuronal spikes records object

and boundary information of image. Outputs of the model are the

desired segmented image after connection components label.

Connected components labeling is implemented on those spike

images to obtain regions and contours through fusing all the

object information recorded in the spike images.

Considering a digital image as network, the network can

achieve image segmentation through encoding the ﬁring rate of

similar neighboring neurons because of the local coupling among

neighbors and decaying exponential. The method ensures that the

results meet segmented image coding based on uniformity within

a region and effectivity of contour pixels.

In the model, an activated neuron may result in the synchro-

nous excitation of its neighbors with approximate intensity. So,

similar neighbors affect each other and could capture local

information. It is called as spiking coupling. Another, the transi-

tion between stimulus is sharpened by gradient coupling and it is

easy to ﬁnd the discontinuous. In short, the model is convenient

for showing the local discontinuous and smoothness of the

network.

The organization of this paper is as follows: In Section 2,

gradient coupling spiking cortex model and its properties for

image processing is demonstrated. Section 3 is the segmentation

method based on the proposed model. The experimental results

will be shown in Sections 4 and 5 is conclusion.

2. Gradient-coupled spiking cortex model and its properties

Inspired by ﬁring rate encoding of neuron network and the

threshold segmentation method, a novel model is constructed to

ﬁnd the discontinuous of stimuli. In this section, we formulate the

model and its properties when it is applied to image processing

especially segmentation.

2.1. Gradient-coupled spiking cortex model

Based on the visual cortical model [22–24], gradient-coupled

spiking cortex model is proposed as follows.

The fundamental component of the model also is leaky integrator.

The basic form of the response is

IðtÞ¼Ve

t=

where V is the ampliﬁcation factor,

is the decay time constant of

the leaky integrator and I(t) is of exponential decay with time t.

Each neuron is denoted with indices (i, j), and its neighbors are

denoted with indices (k, l). Feeding and linking are combined

together as internal activity. Neuron receives input signals from

the stimuli and feedback synapse of its neighbors such that the

output signal of a neuron modulates the activity of its neighbors.

The internal activity U

[n] of neuron in the model is modulated

nonlinearly by feeding input and linking input as

½n¼U

½n1e

t=

þS

½nþS

½n

ijkl

½n1þS

½nGRAD

½n

ð1Þ

R.C. Zhao, Y.D. Ma / Neurocomputing 85 (2012) 45–5246

剩余7页未读，继续阅读

评论收藏

内容反馈

weixin_38681147

粉丝: 7
资源: 937

面向区域图像压缩的区域分割方法

基于区域检测的小波分形图像压缩方法 (2012年)

基于尖峰皮质模型的基于区域的图像压缩新方法

matlab.rar_matlab区域生长_region growing_区域分割_基于区域的区域生长图像分割方法

面向区域分割的车牌定位方法

图像分割毕业论文.doc

LBG矢量量化图像压缩 c++代码

图像压缩算法简介

面向多颜色空间的自适应SFCM聚类图像融合分割.pdf

基于C++ YUV色域转换和LZW编码的BMP图像压缩程序【100012789】

DCT压缩图片整个流程实现

基于安卓平台的作物病害图像自动分割算法实现.pdf

数字图像处理习题大纲

图像分类卷积神经网络的特征选择模型压缩方法（英文）.pdf

数字图像处理课程设计，综合运用C++语言实现图像处理的程序设计

数字图像处理-Java代码

基于MATLAB面向课题的数字图像处理实践教学.zip

Visual C++ 数字图像获取、处理及实践应用随书光盘

C++图像处理，学习数字图像处理的一本好书！

图像处理和分析基础-章毓晋

图像识别技术和图像处理技术.pdf

数字图像处理例程

数字图像处理编程入门

VS数字图像处理

数字图像处理课程综述（英文）

数字图像处理系统

光电图像处理教学大纲.pdf

冈萨雷斯 数字图像处理 英文版答案

最新资源

冈萨雷斯数字图像处理英文版答案