文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 工程科技 > 信息与通信 > A New Image Enhancement Based on the Fuzzy C-Means Clustering
侵权投诉

A New Image Enhancement Based on the Fuzzy C-Means Clustering

gray. Here are the specific steps for conducting clustering in all areas of the infrared dim and small target image clustering.

(1)Transform M × N image into a one-dimensional array, establish a data set X ={x1, x2, …, xn}∈Rn, here, n = N × M, xi is the gray value of the i-th pixel in the image.

(2) The histogram equalization processing of the image [16]: Assume it can be divided into np levels. The i-th level

is used: ifuik max{uik}, then the sample xk Belongs to shares the gray-level ratio of Tj/256, Tj is for the gray of the

01 i c

i-th level, j∈np.

the i0-th class. Therefore, in Practical application, the (3) Calculate the average the gray value of the whole formula (4) can also be expressed as follows: If the number image: of the different points in the membership degree sets of the ncalculation of the two time adjacent is less than a certain

fr

value, then the iteration ends. (5) tavg r 1Improved fuzzy C-means clustering algorithm M NClustering Analysis on the Infrared Dim and Small

Target Image:The key of the infrared image enhancement Here, f is the gray of the i-th pixel.

r

is to increase the gradients of the junctions of the target and

(4) Clustering initialization

background gradients. Use the C-means clustering method

The image data was carried out the clustering

to conduct the gray classification of the image[5], Re-processing, classification number C=np, the initial center

assignment the gray value of each class so as to enhance

the gray difference between various classes. In the ideal value of each class V= {T1, T2,…, Tnp }, the pixel number case, the target should be divided into one class [6, 7]. of each class is rj. Thus, by hierarchical clustering, the data However, it is almost impossible for the dim and small set X whose sample size is n was divided into np subsets, target. Generally speaking, the gray-scale of the dim and npsmall target accounts for fewer pixels, gray difference

that is, S st.

between the target pixel and surrounding other pixel is

t 1

minimal. Therefore, it is easy to class target pixel and

surrounding other pixels of the similar gray into the same (5) Calculate the cluster distance

Since the gray of the pixel points within the each subset class so as not to achieve the purpose of enhancing the

Sk to point Ti can be target. To achieve the classification of the dim and small is closer, so the distance from xi∈

target classification, the only way is to increase the number approximately expressed by the distance from the center Tk

Sk to Ti, namely: of the classes [8, 9]. However, to increase the number of of the

categories can make the clustering speed slow down. In 22

dik(xk,Ti) dik(Tk,Ti) ||Tk Ti||2 addition, due to the presence of noises, even if the number (6)

of categories was raised higher, the noises can only be divided into different classes, the target is difficult to be (6) Calculate the membership, updating cluster centers separated from them[10]. This is because the noise and the The size of the fuzzy matrix U changes from the original background and objectives are very different from the target n×c into n×np. The calculation formula of the membership in the gray levels. Clearly, to directly conduct the gray is still the formula (3), the calculation formula of the np clustering method for the and small target is not feasible. In cluster centers changes from the formula (4) into the follow this paper, the standard C-means clustering method was the formula (7):

improved. Firstly, the infrared image was conducted the

histogram equalization so that the overall gray of the image 4 J r22

J Asin t B12 C2should be evenly distributed. Secondly, the peak value of (7) r3dv 00Vthe histograms was used as the cluster center value. The 0

peak number was used as the number of classes[11]. The

The calculation formula of the objective function gray images were carried out the cluster. Then, the gray

value of the cluster centers were compared and were changes from the formula (1) into the follow the formula (8): analyzed. As the general target area share more pixels and nnp

2the gray difference between the goal and background for (8) Jm U,V (uik)mdik(Tk,Ti)

these images is smaller, therefore, if the pixel number of the k 1i 1certain class is less and its central gray has larger difference from the central gray of other classes, then this (7) Select the appropriate matrix norm (e.g. Euclidean class does not contain the target, whereby it can be distance), compared Jm(k) with Jm(k+1), if ∥Jm(k+1)- Jm(k)∥<ε, removed. The remaining classes will be merged. Then then stop iteration; otherwise k=k+1, go to step (5). according to the original classification number[12], the re-(8) Compare the cluster center value of each class. If

clustering was carried out, and so on, until the difference the cluster center value of the certain class is Tf, Tf >Tj or Tf between each class and other classes in the gray of the

< Tj and the number of pixels contained in the class is rf, rf

cluster centers is little. Even if the dim and small target is

mistakenly divided into several categories, the difference < rj, then Tf=tavg. Here, j =1, 2, …, m, Tj is for the gray value between the target class and other classes in the gray is not of any class, tavg is for the gray value of the entire image.

Therefore, the way that the pixel gray value of the class Tf

large [13, 14], so the target class can not be excluded.

Algorithm Steps:A gray image is usually has a certain to be eliminated is set into the average gray value of the density distribution around the certain number center points whole image can make the class removed.

(9) n1 times continuously perform the step (4). If the

in the gray space [15]. The points located in the vicinity of

noise of the image is larger, n1 value to be taken.

the center point constitute a subset which is similar in the

When the algorithm converges, we get various the centers of the various classes and membership values of various samples belonging to various classes, thus completing the division of fuzzy clustering. If necessary, you can also conduct to remove fuzzy of the results of fuzzy clustering, that is, you can use certain rules divides the fuzzy clustering classification into the deterministic classification. Generally the maximum membership principle

2 PRZEGL D ELEKTROTECHNICZNY (Electrical Review), ISSN 0033-2097, R. 88 NR 3b/2012

上一页第2页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email