基于灰度信息直方图的X射线图像对比度增强改进

所需积分/C币:50 2018-07-08 11:41:56 2.84MB PDF
87
收藏 收藏
举报

用于图像增强目的的基于直方图的技术的许多应用是众所周知的。然而,这些技术对于各种低对比度图像(例如,X射线图像)通常不能产生令人满意的结果。在本文中,我们提出了一种新的图像对比度增强直方图。首先根据梯度强度将输入图像分成几个相同大小的区域,然后分别修改相应的灰度统计值,最后通过所有加权值的总和得到整幅图像的处理直方图的地区。这种新的直方图形式的基本特征是其分量的幅度可以客观地反映灰度级对图像信息表示的贡献。因此,这个新的直方图被称为灰度信息直方图。使用所提出的直方图可以显着改善许多基于直方图的增强技术的性能。在X射线图像上进行测试可验证新直方图的有效性。
M. Zeng et aL. Optik 123(2012)511-520 513 a h 12000r 5000 4000 dxn右E三z 3000 2000 1000 100 Gray Leve Gray Level Gray Level 3000 1200 2500 1000 600 贝x 800 400 600 300 日1000 400 200 mli 100150200250 50100150200250 100 150 200250 Gray Level Gray Level Fig. 2. Comparison of the GHE results using different local histogr dms of regionS. (d Low-contrast original image, and (c) its corresponding standard histogram, (b)GHE result using standard histogram. Five regions marked with white pixels: ld)region1. (g)region2 (j)region3. (m)region4. (p)region5. their corresponding local histograms (,(i),(,(o(r), and their GHE resullts:(e), (h),(k). (n)and (q) 514 M. Zeng et aL. /Optik 123(2012)511-520 AO Al A2 A7 B(x,y) A3 A6 A5 A4 Fig 4. Numbering convention for gradient calculation. where GR(x, y)and gc(x, y) are the row gradient and the column gradient at(x, y), respectively GR(x,y)=n[02+2A3+A1)-(Ao+27+A6) b Gc(x, y)=,[(Ao +2A1+A2)-(A6+2A5 +A4) (3) (3)Divide the original image into a proper number of equal-sized regions(i.e, default five regions), according to ascending orde of their intensities of gradients. Fig. 2 shows an example of this segmentation operation. Each pixel location in the original image has a specific statistical weighting coefficient depending on its gradient magnitude. The coefficients are determined as W1jfG(X,y)≤Tt W2 if T1<Gx,y)≤T2 W(x, y)= w3 if T2< G(x, y)s T3 W4f73<G(x,y)≤74 Ws otherwise where Ti(i=1, 2, 3, 4) denotes the limits of the gradient intervals, Wi(i=1, 2, 3, 4, 5) is the statistical weighting coef- ficients( usually w1≤w2≤W3≤W4≤ Ws due to the different contribution to image contrast enhancement the next half of Section 2 will discuss how to determine the optimal weighting coefficients) (4)Compute the cumulative summation of the weighted statistical values of each gray level in different regions N(r) Wini (r) i=1 n (r)is the number of pixels at the gray level r in a certain region N(r)is the summation of all weighted statistical values of the gray level r in the five regions. It should be pointed out that the standard histogram is a special case of our new histogram that is, when W1=W2=W3=W4=W5= 1, the components of the two types of histograms have the same values. 2.2. Determination of the optimal parameters In the previous paragraph, we describe the basic procedure of this new histogram In order to obtain desired histogram shape for Fig 3. Test example. (a)Original Lena image, (b)result of GHE using local histogram of region1 and(c)result of GHE using local histogram of region5 image contrast enhancement, several problems should be solved Here, we list these problems as follows (2)Calculate the gradient magnitude at each pixel using equations Choose a proper number of equal-sized regions similar to the Sobel edge detector. a 3 x 3 pixel gradient oper Choose a robust method to evaluate the distribution of detailed ator is described by the pixel numbering convention of Fig 4 information The square root gradient is defined as Choose the optimal weighting coefficients G(x, y)=IGR(x, y)+[Gc(x, y)]21 As mentioned previously, we use a default value(i.e, 5)as the total number of regions the choice of this number is also based on M. Zeng et aL. Optik 123(2012)511-520 515 k D h Fig. 5. Test examples. (a)(d)original images, and(e)(h) the corresponding edge detection results, (i)(I)Segmentation results of region5 for(a)(d) the fact that the percentage area of the detailed regions in the entire not very sensitive to the number of the regions if the number is over image is usually in the range of 20-80%(e.g, the detailed region 4, and a large number of regions would cost more computationally covers nearly 20% area in the simple line drawing(Fig. 5(a)), but The next problem is to develop a robust method to evaluate the over% in the complex image of a town, as shown in Fig. 5(d)). We distribution of details in a given image. This is important because choose the lowest limit(i.e, 20%)as the size of the region, and there- it can provide useful information for estimating the regians'con- fore the corresponding total number of regions is five. Moreover, tribution for the image enhancement. Here we propose a simple from multiple trials, we found that the resulting histogram shape is solution. We use the Canny edge detection technique to evaluate 516 M. Zeng et aL. /Optik 123(2012)511-520 a a Fig. 7. Zoomed images for(d)Fig. 5(c). (b) Fig. 5(g),(c)Fig. 5(k where ni is the number of edge points in region i, Ns denotes the number of edge points in region5 which contains the biggest num- ber of edge points Fig. 6. Zoomed images for (a) Fig. 5(b), (b)Fig. 5(1), (c)Fig. 5() Using Eq (6), normalized weighting coefficients for Fig. 5(a)-(d), the distribution of details. The Canny edge detection algorithm is are calculated and listed in Table 2. Here, we highlight some exam ples to show the effectiveness of this normalized operation For a robust and accurate edge detection methods. The motivation for the simple images, e.g., line drawing in Fig. 5(a), region5 with the this solution is that if a region contains many details, it generally has large number of edge points. In other words, the amount of details highest gradients covers all the details, as shown in Fig. 5(i),So is equivalent to the number of edge points. Therefore, the distribu- that the weighting coefficients of other regions are zero which tion of details in different regions is executed as follows: (1)detect dicates these regions have no details. For more complex images, the edge in the input image using the Canny edge detection oper- the region5 does not capture all details(see zoomed images of ator;(2) count the number of edge points in different regions. Test Figs 6 and 7), e.g., in Fig. 7, the minor textures on the cap are not examples are given inFig. 5. The results of edge detection are shown included in the region5. Some details are located in region4 and in Fig. 5(e)(h). The number of edge points in different regions is region3, and the corresponding coefficients show their contribu- listed in table 1 tion to the representation of image details. It is worth noting that The final problem is how to choose appropriate weighing coef- these results agree well with those of visual perception ficients in order to perform the weighting process to yield a desired histogram shape for image contrast enhancement. One can specify 3. Experimental results and comparisons these coefficients in an empirical way to yield satisfactory results but this precludes the new histogram from being applied in many To verify the effectiveness of our new form of histogram, cases in which full automatic processing are needed. Furthermore we apply this new histogram to two conventional histogram since there are five parameters, direct trial-and-error requires a lot based image enhancement techniques: GHE method (a well-known of work. Here, we developed a simple method to overcome this global technique) and CLahe method (a popular local adaptive problem. Our idea is to normalize the number of edge points in technique). Incorporation of gray-level information histogram into different regions to generate their corresponding weighting coeffi the existing contrast enhancement techniques is straightforward cients. The equation can be represented as follows One only needs to insert the new histogram into these algorithms 1,2,3,4,5 (6 In order to distinguish the enhancement methods using gray-level information histogram from the conventional methods using stan M. Zeng et aL. Optik 123(2012)511-520 517 3.5 ≌x6oE 0100150200250 Gray Level 2.5 1.5 oxn.ooE2z 0.5 0.5 200 250 Gray Level Gray Level 2510 10 o×.Ez 0.5 150 200 50 100150200 250 Gray Level Gray Level Fig. 8. Comparison of the results for the line drawing using different enhancement techniques. (a)Low-contrast original image, and (b)its standard histogram, (c)Gray-level information histogram. Note that the amplitude of the rightmost component has been reduced. (d)Result of giie, (e)result of CLAlIe, and (f)result of optimized GIIe; (g) hiistugr'am ford,(I Histogram for (e), and (i llistogrdn for(D dard histogram, we define the method using the new histogram as tains six nonzero components Due to the highest amplitude of the optimized method, e.g., optimized ghe method rightmost component, the resulting histogram of GHE is shifted Fig. 8 shows the optimized ghe result of the line drawing with toward the left side of the grayscale and most of the grayscale comparison to the ghe result and clahe result. Fig 8(a)illustrates empty, which causes significant contrast loss of detailed regions a line drawing, and Fig 8(b) its standard histogram which con- as shown in Fig 8(d). Note that the contrast of ghe result is even Table l number of the edge points in different regions Region1 Region2 Region3 Region4 5 Line drawing 0 0 12,150 Mammography 3216 7031 Lena 0000 12,114 Town 2507 5078 7334 Table 2 Normalized weighting coefficients for different regions W3 Line drawing Mammography 0 0 0.003 0.46 Lena 0.006 0.18 0.57 Town 0.11 0.34 C.69 518 M. Zeng et aL. /Optik 123(2012)511-520 a d Fig. 9. Comparison of the results for X ray image of two hands using different enhancement techniques. (a) Original image, (b) result of GHE, (c)result of ClaHe, (d result of optimized GIIE, and (e)result of optimized CLAIIe worse than that the original image. Fig 8(e)is the Clahe result. It area in the original image and gray levels in this area are con can be seen that although CLAHE method can produce strong con- centrated into a narrower range of grayscale(i.e, from 75 to 100 trastenhancement, the characteror content of the image(e. g pixel grayscale interval, as shown in Fig. 2 (c)), which causes significant value order or the number of nonzero components )is changed due contrast loss of small detailed regions. Moreover, it can be seen to the local mapping nature of this technique, which is undesirable that noise in the background is almost invisible before enhance for many applications. Fig 5(i shows that region5 covers all the ment. However, it becomes apparent after enhancement Fig 9(d details in the line drawing, and the corresponding weighting coef- shows the enhanced X-ray image after applying our ficients are W,=W2=W3=W4=0, W5=1, which means that togram into the Ghe technique. It can be seen that the result of the the components of the resulting histogram in Fig. 8( c) is only the optimized ghe technique is superior to that of the conventional frequency of gray levels in region5. Note that the amplitude of Ghe algorithm. The enhanced image, Fig. 9(d), not only exposes the rightmost component has been reduced Fig. 8(f shows the more image details, but also has little noise overenhancement enhanced image using optimized ghe method, and Fig 8(1)its his- Fig 9(c)shows CLAHE result, which is better than the ghe result togram in which the six components are uniformly spread over the(Fig 9(h), but still has an undesirable appearance(i.e, noise overen- entire grayscale. It is obvious that the contrast of the optimized hancement Fig 9(e) is the optimized clahe result. It is obvious GHE result is better than those of the CLAHE result and the original that the optimized Clahe algorithm outperforms the Clahe and image GHE techniques. Note that there is not much difference between Fig 9 shows the comparison of the results of treating the x-ray the optimized clahe result and ghe result, and both images are image of two hands with different enhancement techniques includ atisfactory. ing GHE, Optimized GHe, CLAHE and Optimized clahe Fig g(a) From empirical results, we found that using our new histogram, shows a low-contrast X-ray image of two hands. Its GHE result, even the simple method(e.g, GHE) can be conveniently applied to Fig 9(b), also looks faded and background noise has been ampli- a broad variety of low-contrast images(e.g, simple line drawing fied. as mentioned in section 2. the cause for this unsatisfactory images with strong noise, etc. ) and yields satisfactory results. More appearance is that the background occupies nearly 60 percent of examples are given in Figs. 10 and 1 M. Zeng et aL. Optik 123(2012)511-520 519 a b Fig. 10. Test example. (a)Original image of hand, (b)result of Ghe and(c)result of Fig. 11. Test example. (a)original image of mammography (b) result of GHe and optimized GHE (c]result of optimiz.ed GHE 4. Conclusion results. Therefore, this paper introduces a new form of histogram called gray-level information histogram. Compared with the con- It has been found that the regions with more details make great ventional histogram, the amplitudes of its components can provide contribution to the image enhancement. However the components accurate assessment of the contribution of different gray levels of the standard histogram only show the frequency of gray levels, to the image details depicting. Testing on various kinds of x-ray while ignoring the distribution of the details. This makes it diffi- images shows that after using our new histogram, the optimized cult for many histogram-based techniques to generate satisfactory histogram-based techniques usually outperform the original meth 520 M. Zeng et aL. /Optik 123(2012)511-520 ods with the same processing parameters. In addition, it should be [9] S.D. Chen, A.R. Ramli. Minimum mean brightness error bi-histogIalm equal pointed out that this proposed histogram is a new representation ization in contrast enhancement, IEEE Trans. Consum. Electron. 49(2003) of the image. Like the standard histogram, it can be applied to other [101 D. Menotti. L Naiman. J Facon. A.A. Araujo. Multi-histogram equalization meth- image processing applications, e. g content-based image retrieval ods for contrast enhancement and brightness preserving, IEEE Trans. Consum. which is currently under study in our research Electron532007)1186-1194 [11] N.S. P Kong, H D. Ibrahim, Color image enhancement using brightness preserv- ing dynamic histogram equalization, IEEE Trdnls ConsuIl Electron. 54(2008) Acknowledgements 1962-1968 [12]G H Park, HH Cho. M R. Choi, a contrast enhancement method using dynamic This work was supported in part by the National Natural sci- range separate histogram equalization. IEEE Trans. Consum Electron. 54(4) (2008)1981-1987 ence Foundation of China under Grant Nos. 60802051, 60875053, [13 T. Kim, J. Paik, Adaptive contrast enhancement using gain-controllable in part by Tianjin Natural Science Foundation under Grant No clipped histogram equalization, IEEE Trans. Consum Electron. 54 (4)(2008) 09JCYBJCO2100, in part by Youth Foundation of Tianjin University 1803-1810. and in part by China 863 High-Tech Program under grant No [14]RA. Hummel. Image enhancement by histogram transformation, Comput Graph Image Process. 6(1977)184-195 2007AA04Z219. M. Zeng would like to thank the support of City [151 S.M. Pizer, E.P. Amburn. J.D. Austin, R. Cromartie. A. Geselowitz. T Greer University of Hong Kong( Project No. 7002511) B H. Romeny, J B. Zimmerman, K. Zuiderveld. Adaptive histogram equal ization and its variations, Comput. Vis. Graph. Image Process. 39(1987) 255-368. References [16]Y P Jin. L Fayad. A Laine. Contrast enhancement by multi-scale adaptive his togram equalization. Proc. SPIE 4478(2001)206-213 [1 A LaSso, E Trucco, Vesselenllancement in digitalX-I'dy angiographic sequences [17I JA. Stark, WJ. Fitzgerald An alternative algorithm for adaptive histogram by temporal statistical learning, Comput. Med. ImAging Graph. 29(2005) equalization Graph Models Image Process. 58(1996)180-185 343-355 [18]R B Paranjape. W M. Morrow, R.M. Rangayyan, Adaptive-neighborhood his [2 G D Wang, N Sang L.X. Yan. X B Shen, X-ray angiogram images enhancement togram equalization for image enhancement Graph. Models Image Process. 54 by facet-based adaptive anisotropic diffusion, Comput. Med. Imaging graph 1992)259-267 33(2009)140-147 [19] J.A. Stark, Adaptive image contrast enhancement using generalizations of his- [3 S.C. Yang, C M. Wang, HH. Hsu, P C Chung, G C. Hsu, C I Juan, C.S. Lo, Contrast togram equalization, IEF Trans. Image Process. 9(2000)889-896 enhancement and tissues classi fication of breast MRI using Kalman filter-baSed [201 M.A.A. Wadud, M.H. Kabir, M.A.A. Dewan. O Chae. a dynamic histogram equal- linedr nixing method, Comput. Med Imaging Graph. 33(2009)187-196 ization for image contrast enhancement, IEEE Trans. Consum. Electron. 53 [4R. H. Sherrier, G.A. Johnson, Regionally adaptive histogI'dm equalization of the (2007)593-600 chest, IEEE Trans. Med. Inaging MI-6(1987)1-7. [211 G. Finlayson, S Hordley, G. Schaefer, G.Y. Tian Illuminant and device invariant RV colour using histogram equalization, Pattern Recogn. 38(2005)179-190 Heercum. Chest cT window settings with multiscale adaptive histogram equal- 122 V. Caselles. J. L. Lisani. J M Morel G Sapiro, Shape preserving local histogram tion: pilot study, Radiology 223(2002)845-852 [6] S.M. Pizer, J B Zimmerman. E.V. Staab, Adaptive grey level assignment in CT [23] T.K. Kim, J.K. Paik, B.S. Kang, Contrast enhancement system using spatially scan display, J. Comput. Assist. Tomogr 8(1984)300-305 adaptive histogram equaliz. ation with temporal filtering, IEFF Trans. Consum [7I S.S. Agaian. B Silver, K A Panetta Transform coefficient histogram-based image Electron.44、1998)82-86. enhancement algorithms using contrast entropy. IEEE Trans. Image Process. 16 [24 J.Y. Kim, L.S. Kim, S H Wang, An advanced contrast enhancement using partially 2007)741-758 overlapped suh-block histogram equalization, IFFF. Trans. Circuits Syst. vide [8 Y.T. Kim Enhancement using brightness preserving bi-histogram equalization. Technol.11(2001475-477 IEEE Trans. Consum. Electron. 43(1997)1-8

...展开详情
试读 10P 基于灰度信息直方图的X射线图像对比度增强改进
立即下载 身份认证后 购VIP低至7折
一个资源只可评论一次,评论内容不能少于5个字
您会向同学/朋友/同事推荐我们的CSDN下载吗?
谢谢参与!您的真实评价是我们改进的动力~
  • 签到新秀

  • 分享宗师

关注 私信
上传资源赚钱or赚积分
最新推荐
基于灰度信息直方图的X射线图像对比度增强改进 50积分/C币 立即下载
1/10
基于灰度信息直方图的X射线图像对比度增强改进第1页
基于灰度信息直方图的X射线图像对比度增强改进第2页

试读结束, 可继续读1页

50积分/C币 立即下载