数字图像处理毕业论文外文翻译

. Gross approximation of the low threshold The low threshold for hysteresis is very important in negative detection because these edge points will not change its state during the linkage stage of the hysteresis process. Accuracy refers to the probability that a pixel will be properly classified (regardless of whether it is positive or negative). Therefore, our proposal should consider to make maximum Ac(x) for the approximation of the low , for the same objective, it is important to detect enough number of positives, making minimum the number of false positives (FP),because these pixels can be re-classified during the linkage stage of the hysteresis process. Then our proposal should consider to make maximum Pr(x) also. Thus we propose the function Low with real valuesLow : {1, . . . , L 1} RLow(x) = Pr(x) + Ac(x), x ∈ {1, . . . , L 1} to characterise the edge map Gb (I) in terms of the probability that a pixel will be accurately classified (Ac(x)) and considering the probability that a pixel will be a true positive (Pr(x)). The maximum x0 of the Low(x) that satisfies H(x0 ) 0 will be the gross low threshold proposed. Given that 1= TP(x) + FP(x) and that 1 = TN(x) + FN(x) it can be written that x ∈ {1, . . . , L 1} Pr(x) = (1/ )TP(x) and Ac(x) = TP(x) +TN(x).. Gross approximation of the high threshold The high threshold is very important in positive detection because these pixels will not change its state during the linkage stage of the hysteresis process. Then we should look for the high threshold to make minimum the number of FP. It could be thought that the number of false negatives (FN) would have a smaller importance todetermine the approximate high threshold because pixels labeled as negatives could change their state during the linkage stage of the hysteresis process. It is our opinion that the above reasoning is not right in general because an initial high number of FN could makemore difficult the linkage stage of the hysteresis process. Therefore, with lack of prior knowledge, our proposal is to find the edge map Gb (I) for which the number of FN and FP are the same. xThis property means selecting an edge map in which the probabilities of error are equal for positives and negatives. The value x0 | FN(x0 ) =FP(x0 ) will be our proposal for the gross approximation of the high threshold. Then to make the gross approximation of the high threshold it is needed to calculate FN(x) and FP(x), x ∈ {1, . . . , L 1}. The other alternative is to use value to determine this threshold. Since value has been calculated above (see Section ), and to calculate value only an x value is necessary, using this alternative, the computing time to calculate this threshold using the 1 percentile of the histogram of the feature image GH (I) could be less. Notice that if for a specific application some prior information is known (for example suppose FP(x0 ) = kFN(x0 )) a similar reasoning is possible to determine the high threshold corresponding to this prior . Fine approximation of hysteresis thresholds Now we are interested in finding a criterion to narrow the search for the hysteresis thresholds. This will be achieved by confining oursearch to the interior of the interval determined by the gross approximation proposed in Section . Fig. 3 shows a graph of the histograms corresponding to the precision, accuracy, sensitivity and specificity (see Eqs. (9)–(12)) obtained for the airplane gradient image. The edge maps determined using the gross approximation of the high and low thresholds are L possible edge maps are represented on the abscissas axis with L = 100. Given the basic principle upon which hysteresis is founded, the edge maps Gb (I) and Gb (I) should have different properties. low high The negatives in Gb (I) can change their state during the linkage high stage of hysteresis. Since specificity is the probability that a pixel that is negative in the reference edge map will be correctly classified as being negative (true negative) our proposal should use specificitySp. On the other hand, the positives in Gb (I) can change their state lowduring the linkage stage of hysteresis. Sensitivity is the probability that a pixel that is positive in the reference edge map will be correctly classified as positive (true positive). Due to this property our proposal should use sensitivity Sn. Therefore, we propose shape sharp changes of sensitivity and specificity histograms must be considered like information to determine the candidates for the hysteresis thresholds. The local extremes of the curvature is a well-known technique to characterise the shape histogram [5,6]. We use this technique to determine significant bins in the sensitivity and specificity histograms and our proposal is to use these bins as the candidates for the hysteresis thresholds. For example, Fig. 4(a) shows the graphs of Sn(x) and Sp(x) for the airplane gradient image in the gross approximation . 4(b) shows the graphs corresponding to 2 Sn(x) and 2 Sp(x).The local extremes of the second difference correspond to the values of Sn and Sp in which the curvature is maximum or minimum. We know this supposition implies to assume an error. Two different errors are accumulated by our proposal: Our supposition that Pr(x), Ac(x), Sn(x), Sp(x) are approximate values of the true values Pr(x), Ac(x), Sn(x), Sp(x). The choice of local extremes as the candidates for the hysteresis accumulated error will be evaluated with our experimentation.

Zhu W , , Multiresolution watermarking for image and video 1999(04)2. Wang H-J , , Wavelet-Based digital image watermarking 1998(12)3. Swanson , , A Multimedia data embedding and watermarking technologies [外文期刊] 1998(06)4. Wolfgang , , E Perpetual watermarks for digital images and video 1999(07)5. Delaigle , B Watermarking algorithm based on a human visual model [外文期刊] 1998(03)6. Podilchuk , W Image-Adaptive watermarking using visual models [外文期刊] 1998(04)7. Shapiro J M Embedded image coding using zerotrees of wavelet coefficients [外文期刊] 1993(12)8. Zou H Discrete orthogonal M-Band wavelet decompositions [外文会议] 19929. Mallat , S Characterization of signals form multiscale edges 1992(07)10. Cox , , T Secure spread spectrum watermarking for multimedia [外文期刊] 1997

你需要图像处理什么方向的？ 找几篇硕博士论文的摘要抄吧，或者找到这方面教程的中英文版本，呵呵！！！！！！！！！！！！

针对给出的图像（）或者自行选择的灰度图像：1）：给图像分别添加高斯噪声和椒盐噪声。2）：对加噪图像的中心区域（100*100）进行空间滤波，尽最大可能消除噪声。3）：对加噪图像的中心区域（100*100）进行频域滤波，尽最大可能消除噪声。技术描述：对图像进行加高斯噪声和椒盐噪声处理；对包含高斯噪声和椒盐噪声的图片进行处理，使处理后的图像比原图像清晰。所需应用到的技术，包括：a>对图片加噪声b>选取中心区域c>对选取的区域进行降噪处理d>重新生成图像。e>构造高斯低通滤波器时用到了高斯公式：exp(-(u^2+v^2)/(2*(D0^2)))结果讨论：以下是对不同的滤波器针对不同噪点处理的测试结果。参考下面的试验结果，进行讨论：A(011)是使用fspecial('gaussian’)平滑空域滤波处理效果，不过效果不是最好，由于最大程度降噪，导致图像模糊；A(012)是频域滤波处理后的结果，因为使用了高斯低通滤波器，所以会有条黑线，处理一般；A(021)是使用medfilt2（）空域中值滤波器效果，降噪效果很不错，图像也很清晰；A(022) 是频域滤波处理后的结果，同A(012)，因为使用了高斯低通滤波器，所以会有条黑线，效果一般。试验结果:高斯加噪和椒盐加噪处理图分别如下:如图:图(A00):原图图(A01):高斯加噪图(A011):对图(A01)进行中心100*100空域滤波图(A012):对图(A01)进行中心100*100频域滤波（A0） （A01）（A011） (A012)如图:图(A00):原图图(A02):椒盐加噪图(A021):对图(A02)进行中心100*100空域滤波图(A022):对图(A02)进行中心100*100频域滤波(A00) (A02)(A021) (A022)附录：源代码1 ：对高斯噪声的处理f=imread('');J=imnoise(f,'gaussian',);%添加高斯噪声%空域滤波r=[219 319 319 219];c=[129 129 229 229];BW=roipoly(J,c,r);h=fspecial('gaussian',[5 5]);A011=roifilt2(J,h,BW);figure,imshow(A011);%频域滤波f1=imcrop(fn,[129 219 99 100]);%截取100*100大小的窗口图片f2=[255 255];%建立一个新的图像f2=uint8(f2);f2=padarray(f2,[50 49],255);%将新建图像拓展到100*100的黑色图片f2=padarray(f2,[218 129],0);%在新建图片周围添加白色使之大小为moon图片的大小fn=fn-f2;%得到中心100*100区域内为黑色的moon图片PQ=paddedsize(size(f1));[u,v]=dftuv(PQ(1),PQ(2));D0=*PQ(2);hh=exp(-(u.^2+v.^2)/(2*(D0^2)));%构造高斯低通滤波器h1=dftfilt(f1,hh);A012=padarray(h1,[218 129],0);%将h1拓展到moon图片大小A012=uint8(A012)+fn;%得到中心100*100区域处理后的moon图片figure,imshow(A012);源代码2 ：对椒盐噪声的处理f=imread('');fn=imnoise(f,'salt & pepper',);%添加椒盐噪声%空域滤波f1=imcrop(fn,[129 219 99 100]);%截取100*100大小的窗口图片f2=[255 255];%建立一个新的图像f2=uint8(f2);f2=padarray(f2,[50 49],255);%将新建图像拓展到100*100的黑色图片f2=padarray(f2,[218 129],0);%在新建图片周围添加白色使之大小为moon图片的大小fn=fn-f2;%得到中心100*100区域内为黑色的moon图片h=medfilt2(f1,'symmetric');%对f1进行中值处理A021=padarray(h,[218 129],0);%将h拓展到moon图片大小A021=A021+fn;%得到中心100*100区域处理后的moon图片figure,imshow(A021);%频域滤波f1=imcrop(fn,[129 219 99 100]);%截取100*100大小的窗口图片f2=[255 255];%建立一个新的图像f2=uint8(f2);f2=padarray(f2,[50 49],255);%将新建图像拓展到100*100的黑色图片f2=padarray(f2,[218 129],0);%在新建图片周围添加白色使之大小为moon图片的大小fn=fn-f2;%得到中心100*100区域内为黑色的moon图片PQ=paddedsize(size(f1));[u,v]=dftuv(PQ(1),PQ(2));D0=*PQ(2);hh=exp(-(u.^2+v.^2)/(2*(D0^2)));%构造高斯低通滤波器h1=dftfilt(f1,hh);A022=padarray(h1,[218 129],0);%将h1拓展到moon图片大小A022=uint8(A022)+fn;%得到中心100*100区域处理后的moon图片figure,imshow(A022);