当前位置：文档库 › 2008 Automated Color Image Edge Detection Using PCNN Model

2008 Automated Color Image Edge Detection Using PCNN Model

Automated Color Image Edge Detection Using

Improved PCNN Model

LIANG ZHOU, YU SUN, JIANGUO ZHENG

Glorious Sun School of Business & Management

Donghua University

Shanghai, 200051 China

1882 West Yan’An Rd., Shanghai 200051

P. R. CHINA

homzhou@https://www.wendangku.net/doc/7515920938.html,, sunyu@https://www.wendangku.net/doc/7515920938.html,

Abstract: -Recent researches indicate that pulse coupled neural network can be used for image processing, such as image segmentation and edge detection effectively. However, up to now it has mainly been used for the processing of gray images or binary images, and the parameters of the network are always adjusted and confirmed manually for different images, which impede PCNN’s application in image processing. To solve these problems, based on the model of Pulse Coupled Neural Network and the model of HIS, this paper bring forward an improved PCNN model in the color image segmentation with the parameters determined by images’ spatial and gray characteristics automatically at the first, then use the above model to obtain the edge information. The experiment results show the good effect of the new PCNN model.

Key-Words: - pulse coupled neural network (PCNN), image processing, HIS, color image segmentation, parameter determination, image edge detection, spatial characteristics, gray characteristics

1 Introduction

Pulse Coupled Neural Network, PCNN [1] was brought forward by Eckhorn firstly. It is a model derived from a neural mammal model. PCNN models have biological background and are based on the experimental observations of synchronous pulse bursts in the cat and the monkey visual cortex. Johnson and his colleagues had modified the model of the initial PCNN, which was applicable to calculate in the computer more easily. The improved model of PCNN have the well feature of transmitting burst, which is used widely in the fields of image processing, pattern recognition and so on.

The existed algorithm of PCNN is often applied to the fields of segmentation of gray image[2], edge detection of gray image[3] or binary image[4] respectively, Bao qingfeng put forward a new algorithm based on PCNN firstly[5], in her paper, Bao used PCNN in the colorized image. However, Bao’s algorithm is based on Xiao Dong Gu’s [6], which is a simplified algorithm of image segmentation based on PCNN. In Bao’s algorithm, the parameter of network should be adjusted manually firstly, and then it could be used to segment the image. At the same time, the author didn’t use the model of PCNN to process edge detection of the image. So if she wanted to get the result of segmentation, she should adjust the parameters time after time, and switch the model of PCNN when edge detection began. This algorithm restricted the application of the PCNN in the fields of color image. In order to improve the model of PCNN, we put forward a new PCNN system, which can adjust the parameters of the network voluntarily based on the reference [5], and we can use it to process the segmentation and edge detection of colorized image. The experimental results show that we can get the approving results.

2 PCNN

PCNN is a neural network which is composed of lots of neurons. Every neuron is made up of dendritic tree, linking modulation, and pulse generator. Dendritic tree is divided into two small portions, which are applied to guide two different of input to

the linking part of the neuron. One of them is applied to accept the outside input signal which we called it feed-in input, and the other one is used for receive other neuron’s linking input. We can describe them as follows: ∑+?+??=ij kl ijkl F ij ij I n Y M V n F n F F )1(]1[)exp(][α (1)

∑?+??=]1[]1[)exp(][n Y W V n L n L kl ijkl L ij ij L α (2)

])

[1]([][n L n F n U ij ij ij β+= (3) ?

?>=otherwise n T n U n Y ij ij ij ,0]

[][1][， (4) ]

[]1[)exp(][n Y V n T n T ij T ij ij T +??=α (5) As the top formulas, ，，，，，

mean separately neurons’ outside stimulating input, feed-in input, linking input, inside activity, firing output, dynamic threshold. ij I ij F ij L ij U ij Y ij T M and W are

linking weight matrix (usually, )，V ，， mean separately inherent electricity in ，，and , W M =F L V T V ij F ij L ij T F α，L α，T α mean separately

attenuation time constants in ，，T , is a

loop constant, and is a binary output.

ij F ij L ij n ij Y As shown in figure 1, the neuron receives input signals from other neurons and from external sources through the receptive field. In general, the signals from other neurons are pulses; the signals from external sources are analog timing-varying signals, constants, or pulses. After inputting the receptive field, input signals are divided into two channels. One channel is feed-in input (); the other is linking input (). In general, the feeding connection has a slower characteristic response time constant than that of the linking connection. In modulation field, see Fig.1 and Equ. (3), at first, the linking input is added a constant positive bias. Then it is multiplied by the feed-in input. The bias is taken to be unity, ij F ij L β is the linking strength. The total inside activity is the result of modulation. Because the feed-in input has a slower characteristic response time constant than that of the linking input, is like a spikelike signal riding on an approximate constant. The pulse generator consists of a threshold adjuster, a threshold discriminator and a pulse creator. The threshold changes with the variation of the neuron’s output pulse. When the neuron emits a pulse, it feeds back to increase the threshold. When arises more than , the pulse

creator closes and stop emitting pulses. Then

threshold value drops. When threshold drops less than , the pulse creator opens again and emits pulses, namely fires. If the neuron only emits a pulse when it fires, the threshold discriminator and the pulse former can be replaced by a step function. This model is shown in Fig.1. Meanwhile, Equ.(4) shows the neuron’s output under single output pulse condition and is the output. Connecting the neurons on another, then a PCNN model appears.

ij U ij U ij T ij T ij U ij T ij U ij Y Unlike other neural network, there is no training involved for PCNN, and it is an extremely powerful algorithm that can be applied in a variety of ways to image processing. PCNN has the characteristics of grouping of pixels in term of 2D space similarity or gray similarity, reduces local gray difference of image, and makes up local tiny discontinuous points of image. Thus, the algorithm can remove noise, do segmentation, edge detection and object isolation, etc.

Fig.1 The model neuron of PCNN

When PCNN is used for the discrete image processing, it is a single layer 2D array of laterally linked neurons. Each neuron in the processing layer is directly tied to image pixel or set of neighboring image pixels, which means that the number of neurons is equal to that of pixels of input image, every neuron or pixel has its corresponding pixel or neuron, and every neuron is linked to its corresponding pixel and its neighboring neurons. Neuron iteratively processes signals from these nearby image pixels and linking from nearby neurons to produce a pulse train. Similarities in the input pixels cause the associated neurons to fire in synchrony indicating similar structure or texture. This synchrony of pulses is then used to segment similar structures or textures in the image.

3．The description of the new model

Presently, there are two color models in common use

now: RGB model and HIS model. When we use HIS model, it means hue, saturation, and intensity. HIS model has two advantages. Firstly, intensity has not relations with the color information of the image; secondly, hue and saturation have a close relation with the feeling of human. All of these features let HIS model fit for the image segmentation based on the human’s vision system. In this paper, we use the HIS model. At the same time, we propose a new improved PCNN model which can ensure the parameters adaptively based on the traditional PCNN, and in this paper, we apply the same improved model to the color image segmentation and edge detection.

3.1 Confirming the parameters adaptively

Image segmentation and edge detection play important role in any image processing system. PCNN can be efficiently applied to image segmentation and edge detection. The performance of image segmentation and edge detection based on PCNN depends on the suitable PCNN parameters. However, the PCNN parameters which are suitable for segmenting a particular type of images may not be suitable for segmenting a different type of images. In this paper, we should describe how to get the suitable PCNN parameters to efficiently segment images and detect the edge of images.

Firstly, we should transform a color image into the space of HIS, secondly, we set up a corresponding model of PCNN for the hue, saturation, intensity separately, and thirdly we should set image point: to be unitary, at last, we get the final image:.

),(j i ),(j i f When we segment the image, we should set the right parameters based on the image’s features of gray and space, and when we process edge detection, we use the same new PCNN model, and we just need to ensure the right area(gray value=1) of the binary image which has the corresponding neuron, and the corresponding neuron can fire the impulse alone the shape of the target and transform freely, then we control the dynamic threshold, we can get the whole edge of the target. So when we do the edge detection, we can predigest the parameters. The material methods as follows:

(1) Linking weights ：

ijkl W This parameter means how much the output impulse of the surrounding neuron has effect upon the current neuron. As for the current neuron’s corresponding image point, more small the distance between it and its corresponding image point is, more close the differences between these image

points are, then we can conclude that more strong the degree of effect is.

s g s g ijkl l

k l j k i d l j k i d l j k i d l j k i d W )))1),(),((1()1),(),((1(,∑

+++×+++++×++= (6)

Let ),(l j k i d g ++ means the distance of the

gray value between two corresponding image points , and ),(l j k i d s ++ means the distance of space between two corresponding image points.

),(l j k i d g ++ is

defined ：),(),(l j k i f j i f ++? ),(l j k i d g ++ is

defined ：22l k +； When we do the image segmentation: let

1=e ；When we do the edge detection: we want to make the output impulse of neuron transform freely and quickly, let 0=e 。

(2) Adjusting the step of threshold value r ： Let means the times of iterative in the max n image processing, we want to ensure the threshold value can be through all of the image points. The adjusting step of threshold value is: max 1n r = (7) When we do the edge detection: let 0=r .

(3) inherent electricity ：

T V The inherent electricity in the

T V PCNN is used to judge whether the neuron is firing at a moment, if it is firing, then we should set and let the corresponding threshold increase quickly. When >, the impulse creator can stop firing. In this paper, we have the same no matter what kinds of images, and let: (8)

T V T V ij U T V 100=T V (4) initial dynamic threshold ：ij T α=ij T (9) When we do the image segmentation: Let 1=α; When we do the edge detection: Let 15.0=α. (5) modulating parameters β：

e ij ij M V )(=β (10)

when we do the image segmentation: let 1=e , then we can control the degree of increasing of ,

means that image point which corresponding image points’ gray value variance in the surrounding area, and means that image point which corresponding image points’ gray value mean in the surrounding area, when ij F ij V ),(j i ij M ),(j i 0=ij M , let 2.0=β. More small βand the distributing scope of gray value of image points are, more average the distributing is in the surrounding area, then we can find a little improving of gray value can make with the neuron firing in the ),(j i surrounding area at the same time, whereas, the

distributing is not average, then we need a big improving of gray value can make ),(j i with the neuron firing in the surrounding area at the same time. In a certain extent, we can confirm the integrality of the area of segmentation.

When we do the edge detection, we will predigest β, and let 0=e .

3.2 The threshold segmentation of the image ln S (11) and mean the probability and “0”ou based on the Shannon entropy maximum rule

In the reference [7], the author defines the segmentation methods based on the Shannon entropy maximum rule as follows:

11ln )(S S S S H ??=0

01S 0S of the “1”

in the tput binary image separately. The

formula (11) means that how much the information of statistical mean value of the image points (“1” or “0”) include in a binary image after segmentation. Usually, the more Shannon entropy is, it means the image after segmentation get more information from the initial image, and then the image after segmentation has ample details, and well segmentation effect as a whole. In this paper, we will calculate the 1S and 0S from the output binary image Y by an iterative processing. When 1S and 0S make the H maximum, which means w get a t segmentation binary image:e bes Y .

3.3 The new model of color image hould transform the initial color PCNN mod segmentation

At the first, we s image into HIS space, and then we can get the hue image, saturation image, and intensity image.

We should calculate them by improved el respectively, the improved PCNN model as follows:

ij ij I n F =)( (12)

∑∈?=

p p p ij n Y W n L e )1()( (13)

K is a 3*3 matrix, and it means a area which is surrounding the image point

),(j i ))(1)(()(n L n F n U ij ij ij ij e β+= (14)

)()()(n Y V r n T n T T ij ij +?=?

?>=?=else T U if n T n U step n Y ij

ij ij ij ij ,0,1))()(()( (16) Secondly, we use the PCNN model to calculate

the hue image, saturation image, and intensity image separately as follows:

(1) We should transform the gray value of image point into be unitary, and the initial status of the neurons should be set as follows: ij I 0Re ====Y s U L ，

]50,10[max ∈n K is a 3*3 matrix, and all of its value are 1, is a matrix which saves the result, is a matrix which means the firing status, and is the times of iteration, the rest are calculated by the top formulas automatically.

s Re Y n (2)

∑∈=

p p p

ij Y W

L e

；)1(ij ij ij ij L F U e β+=；

)(ij ij ij T U step Y ?=；

(3)α=T ; (4) ∑∈=

p p p

ij Y W

L e

；

(5)),(j i Y Tem =;)

1(ij ij ij ij L F U e

β+=;

)(ij ij ij T U step Y ?=; is a temporary matrix ；

Tem (6) If Tem j i Y =),( go to (7); ∑∈=

p p p

ij Y W

L e

Else go

to (5); (7)If 1),(=j i Y 1),(Re =j i s ，

、 are the corresponding elem of 、；

),(j i Y ),(Re j i s ents Y s Re (8)Y V r j i T T *),(+?=α

(9)To calculate Shannon en py of the

corresponding , if we find the maximum tro image H n , and the when the times of iterative is corresponding )(n Y is the last binary image.

(10)1?=n n ; If 0≠n , go to (4); Else END ；

(15)

(11)To integrate the three binary images, then

we can get the result, which is the segmentation image:Bin .

3.4 The new model of color image edge detection

Edges are very important to any vision system. An edge may be regarded as a boundary between two dissimilar regions in an image. These may be rent surfaces of the object, or perhaps a boundary between light and shadow f diffe alling on a singl nsmits distance which is as e, which width is 5 image points. e surface. In this paper, as to the binary image: Bin , the edge detection algorithm can be obtained by the same PCNN model.

Firstly, we should let the corresponding image points of the bright area fire impulse, and the corresponding image points of the dark area do not fire impulse in this image, secondly, the bright area will transmit as the same width as an image point, then the edge between the dark area and bright area would be fired. If we control the transmitting distance, we can control the width of result of image edge detection expediently. For example, if the impulse of bright area tra the same width as 5 image points, and then we can get a result of imag The model is as follows: (1) To set a matrix: E , which can save the result of edge detection.

should in y ich are 0.1(corresponding to the dark area of Bin ), and 1(corresponding to the bright a We transform the to be unitar wh of Bin rea Bin ),Bin F =,0===U =E Y L , 1max +=N n , is the width of the edge image, and the others are calculated by the top formulas automatically;

(2)

N ∑∈=

p p p

Y W

L e

；)1(L F U e β+=；)(T U step Y ?=；

(3) If , it means:

A neuron is inspired to fire, w neuron’s threshold to stop it fire.

1=ij Y ,ij T ij ij Y V T T += e should increase this (4)if 1+=N n ,1?=n n ,go back to (2)；else

1?=n n

(5)if 1=ij Y ,1=ij E 。

(6)if 0=n ,we get the ou E tput: ，it is the to the image, our results have detected the whole targets’ edge smoothly, and we can also change the width of the last results of images according to the necessary to make these images more clearly.

edge image. Else go back to (2).

4 Experimental Results

We use the new model to process two different kinds of images: figure2 (a) and (b). We can find if we adopt the Bao’s PCNN model, the results are figure 3(a) and (c) which create noises, some bright areas are short of segmentation, and the details are not good. When Bao detect image edge, the results are figure3 (b) and (d). They are not clear, and have lots of disconnected points, so we try to use the traditional method of Laplacian of Gaussian method to get the edges in figure4 (a) and (b). We find that edges are affected by noise present in an image clearly, and the processing results are lack of continuity, too. However, we apply the new PCNN model into the same images, and we get the segmentation results which are figure 5(a) and (c), we can find the details of the image keep integrality; and the shape of the image is intact. The results after edge detection are figure 5(b) and (d), we apply the

same PCNN model in

(a) The initial car’s (b)The initial person’s image image

Fig.2 Original image

(a) After segmentation (b) after edge detection

Fig.3 Process results of model in references ]

(a) After edge detection (b) after edge detection

Fig.4 Process results placian of Gaussia

of La n

Method

(a) After segmentation (b) after edge detection

r, th

t we

hoose it based on the experience and ensure it based

aximum rule, so how to get it more

impr

incre

Refer

[1]

V ol.10, No.3, 1999, pp. 480-498.

005, pp.

[3]

g and

[4]

and

[5]

[6], Dao Heng Yu, A

ing, 2002.

]Ma Yide, Dai Ruolan, Li Lian, Automated

image segmentation using pulse coupled neural

networks and images entropy, Journal of China

Institute of Communications, V ol.23, No.1,

2002, pp. 46-51.

Fig.5 Process results of new model

5 Conclusions

In this paper, we create an improved PCNN model,

which can choose the right network parameters

adaptively, and we apply it into different kinds of color

images in the image segmentation and edge detection.

At the different processing stage, the new PCNN model

can adjust the network parameters automatically. After

the experiment, we can get a robust and intact result no

matter how complexity our processing image is in

images and it would be interesting to see that no matter

how small or large an object is in images, the achieved

result is satisfactory according to this new PCNN

model. Howeve is new PCNN model also has a lot

of disadvantages and limitations. Firstly, the new

PCNN model should waste more time to process

images, which will restrict the application of this model

to the environment of real time processing. Secondly,

the parameter max

n is the times of iteration, which is

a pivotal factor in the image segmentation, bu

on entropy m

reasonably is a problem, too. As an open issue, how to

ove the PCNN model, predigest the model, and

ase its efficiency, we need to research more.

ences:

JOHNSON J L, PADGETTML, PCNN Models

and Application, IEEE Transactions on Neural

Networks,

[2]Liu Qing, Ma Yide, Qian Zhibo, Automated

Image Segmentation Using Improved PCNN

Model Based on Cross-entropy, Journal of

Image and Graphics, V ol.10, No.5,2

579-584.

Yang Zhiyong, Zhou Qiyun, Zhou Dingkang,

Gray Image Edge Detection Method Based on

PCNN, Computer Engineerin

Applications, No.21, 2004, pp. 92-94.

Gu Xiaodong, Guo Shide, Yu Daoheng, A New

Approach for Image Edge Detection Using

PCNN, Computer Engineering

Applications, No.16, 2003, pp. 1-3.

Bao Qingfeng, Wang Jicheng, A New Color

Image Segmentation Based on PCNN,

Computer Engineering and Applications, No.27,

2005, pp. 48-50.

Xiao Dong Gu, Shi De Guo

new approach for automated image

segmentation based on unit-linking pcnn,

IEEE:Proceedings for the First International

Conference on Machine Learning and

Cybernetics, Beij

(完整版)六大基本初等函数图像及其性质

六大基本初等函数图像及其性质一、常值函数（也称常数函数）y =C（其中C 为常数）； α

1）当α为正整数时，函数的定义域为区间为),(+∞-∞∈x ，他们的图形都经过原点，并当α>1时在原点处与x 轴相切。且α为奇数时，图形关于原点对称；α为偶数时图形关于y 轴对称； 2）当α为负整数时。函数的定义域为除去x=0的所有实数； 3）当α为正有理数 n m 时，n 为偶数时函数的定义域为（0, +∞），n 为奇数时函数的定义域为（-∞,+∞），函数的图形均经过原点和（1 ,1）； 4）如果m>n 图形于x 轴相切，如果ma ，1≠a )，定义域是R ； [无界函数] 1.指数函数的图象： 2. 1）当1>a 时函数为单调增,当10<

3.（选，补充）指数函数值的大小比较* N ∈a ； a.底数互为倒数的两个指数函数 x a x f =)(， x a x f ? ? ? ??=1)( 的函数图像关于y 轴对称。 b.1.当1>a 时，a 值越大，x a y = 的图像越靠近y 轴； b.2.当10<∈>=n Z n m a a a n m n m (2)) 1,,,0(1 1*>∈>= =- n Z n m a a a a n m n m n m y x f x x x x g ? ? ?=1)(