Localized Image Segmentation and Enhancement for Meteorite Images

Transcription

1 Localized Image Segmetatio ad Ehacemet for Meteorite Images Yufag Bao, PhD Math ad Computer Sciece Departmet Fayetteville State Uiversity, Fayetteville, NC ABSTRACT This paper proposed a image ehacemet ad segmetatio algorithm to localize the segmet of iterest i a series of images obtaied from the Electro Microprobe. There are totally four differet images that are correspodig to the spatial desity distributios of four major chemical elemets iside a meteorite rock surface. The desity distributio of each chemical elemet is show i a gray valued image with a resolutio i the micrometer rage. Our algorithm applied a statistical image ehacemet techique to improve the visualizatio of features about the same segmet of iterest i three images usig oe image as a referece. The three chemical elemets distributios of a localized cluster were ehaced, ad evetually itegrated together ito a sythetic color image to reflect various chemical compoud distributios isider this cluster. Keywords Image ehacemet, segmetatio, meteorite images. 1. INTRODUCTION The methods for aalyzig images ofte varied accordig to the differet features i images acquired from various imagig istrumets. I the case of studyig the chemical compoud i the ureilites, the typical small size (about 1x1 i 2 ) of ureilites used for aalysis has made measuremet extremely difficult for scietists. The measuremet is powered by usig a ew geeratio imagig system called Electro Microprobe. The Electro Microprobe allows scietists to acquire a set of high spatial resolutio gray images. Each shows the cocetratio of oe chemical elemet. I this paper, four images of chemical distributios of a ureilite rock are acquired from the Electro Microprobe. The images represet the spatial desity distributios of four chemical elemets, Magesium (Mg), Iro (Fe), Alumium (A, ad Calcium (Ca). The Field of view (FOV) of each image covers the polished surface of ureilite with a resolutio i a micrometer (mm) rage [1]. The high itesity pixel value i a image shows high chemical cocetratio, while a low itesity value shows a low chemical cocetratio i a pixel. The four chemical elemet images of the ureilite have the followig features. The itesity values of the Mg image are grouped by clusters; hece, the overall image is of bright ad high cotrast as show (Fig. 1). This is cosistet with the existig kowledge of the rich Mg cocetratios iside a ureilite i geeral. The Iro (Fe), Alumium (A, ad Calcium (Ca) images usually were very dark iside each cluster due to their lower cocetratios i geeral, ad the low itesity values shadowed the boudaries betwee adjacet clusters as show i Fig. 2. It is ot practical to study directly the whole spatial cotet of the images while, ofte, oly a small cluster was of iterest. The images acquired from a Electro Microprobe usually are of large size with aroud 1,000 x 1,500 pixels i total for a ureilite roack of size 1x1 i 2. Not oly is the computatioal burde a cocer, but also the rich itesity scales iside a large size image will degrade the view of the cluster of iterest if the whole image is directly studied. Due to its simultaeous acquisitio feature, a pixel i the same locatio i each image usually accouts for the same positio i the rock surface. It is useful to use oe chemical elemet image, such as the Mg image, as a referece, to idetify the segmets of iterest i the rest images to reveal the material compositio i the rock surface. This is useful for aalyzig all chemical elemet distributio images simultaeously i order to study the highly correlated features iside the images. Aother reaso for arrowig ito a small segmet of the images for study is that the rich itesities coexisted i the image makes it difficult to ehace the features usig the existig algorithms. It is ratioal to search for a small cluster of iterest for ehacemet to avoid the complexity with ehacemet of the etire dark image (see fig. 2)

2 I this paper, we first described a iteractive method to select a cluster of iterest i the rock surface by usig the bright ad high cotrast Mg image as a referece image. The locatio of the selected cluster was the mapped ito the distributio images of the other chemical elemets. The low itesity cotrast iside the cluster will be selected for further processig. Secodly, we describe a method to correct the gray values of the low itesity images by costraiig to this small cluster. The iterior pixels iside this cluster oly cout for a fractio of the origial image size. A improved cotrast withi this cluster will be locally maximized. Fially, we proposed to sythesize all the ehaced segmets of three major chemical elemets ito a color image to characterize the chemical compouds to show the microscopic view of a cluster iside the rock surface. Figure 1: The Mg image to be used as a referece. Figure 2: The Fe image to be ehaced. 2. CLUSTER SEGMENTATION Image segmetatio is ofte associated with edge detectio. Ofte, a closed edge formed with coected curves is of iterest, which typically defies the boudary of a segmeted object. However, most of edge detectio methods are simply based o either the gradiet or the zero crossig iformatio ad therefore are ot geometrically orieted. The gradiet iformatio is used for edge detectio by defiig edges as the local directioal maxima of the absolute gradiet magitude computed. ad is represeted by the Sobel edge detector [2] ad Cay Edge Detector [3]. The zerocrossig iformatio is used for edge detectio by computig a secod-order derivative of the image values [9]. The traditio edge detectio methods are capable of detectig the potetial edges faithfully showig all the high costrast locatios i the image [2-3, 7]; however, it is ot suitable whe our goal is to idetify a cluster i the meteorite image. Our iterest is to label oly the cluster of image for aalyzig the chemical compouds iside the cluster, while eglectig some small segmets iside this cluster. Meteorite images are featured by their massive size (Fig. 1) with differet clusters itegrated together. Each cluster was grouped by pixels with similar itesity values while the boudaries are marked with relatively differet itesity values. As we metioed earlier, the Mg image is the best cadidate for edge detectio. This image is used to idetify a cluster of iterest to allow ehacemet [4] to be applied efficietly to this cluster i the other images. Our proposed edge detectio techique is based o the Sobel detector ad modified morphology methods [8]. To begi, we select a rectagle segmet o the referece, Mg, image. Iside the rectagle regio, the histogram of the image is geerated. The mea of the histogram is used as a automatic threshold level for geeratig a biary image where the pixels of the potetial boudaries of a cluster object are marked as 1, ad others are marked as 0. The iitial operatio of this thresholdig usig the Sobel detector fids all the potetial boudaries blidly, which icludes boudaries of very small segmets. Based o this potetial boudary image, B, the shortest Euclidea distace of each pixel to the potetial boudary i

3 image B is calculated, i.e. for each pixel represeted as (i, j), d(i, j) is defied as d( i, j) = if ( ( i, j) ( k, ) 2 ( k, l ) B The boudary image, B, is updated to uify the multiple broke boudaries that belog to the same edge: 1, B( k, = 0, if if d d ( k, ( k, < threshold threshold where the threshold is set for umber 3. Note this threshold is differet from the threshold used for edge detectio. The biary value of 1 i the image, B, idicates the improved boudaries. This updated boudary allows multiple boudaries for the same segmet to be uified together to show the geometrical features. Fially, the largest coected compoet is idetified; To do so, we idetified all the coected compoets i the edge images, ad the defied the size of each coected compoet, O, as M ( o ) = k I( x, The largest compoet to be preserved is defied as O = arg max M ( o Ok This method discarded all the isolated small segmets iitially beig treated as boudaries. I additio, the morphologically operatios, amely, a dilatio followed by a erosio, are performed to close the potetial gaps iside the boudaries, while to esure the boudaries for a cluster are properly coected. I so doig, the major boudaries are idetified. It separates the Mg image ito several clusters that we are iterested i. Oly the iside of the cluster of iterest was selected ad preserved. The major steps for fidig a cluster of iterest for the Mg image is show i Fig ENHANCEMENT Ehacemet of a low cotrast chemical elemet image allows us to visualize the spatial desity distributio of the compouds made from this elemet, ad provides a base for coutig the O k k ) k cocetratios. Whe a image is i low cotrast, the image ehacemet sometimes ivolves deblurrig ad oise removal procedures [5]. The histogram of a image statistically demostrates the the image itesity distributio ad is typically the base for cotrast correctio. Direct cotrast adjustmet methods [2] ad histogram equalizatio methods [6-7] are two methods geerally used for gray image ehacemet. To better preserve the chemical elemet cocetratios, histogram equalizatio methods is used i our study. The histogram equalizatio methods map the iput gray levels to the output gray levels which are evely dispersed based o statistical kowledge. The existig image ehacemet algorithm is improved to ehace the distributio of each chemical elemet image. a) Select a rectagle area from the origial Image. c) The largest coected compoet with ew boudaries defied ad morphological operatio. b) A biary image was geerated by applyig Sobel detectio with a automatic threshold. d) A iteractive selectio of the cluster with closed cotour. Figure 3. Our proposed cluster segmetatio The traditioal histogram equalizatio methods adjust the image itesity globally. It treats the etire image equivalet ad usually yields poor local performace; therefore, i a localized regio, the itesity may ot be appropriately separated eve after ehacemet. Several improvemet algorithms to ehace the itesity i a localized regio have sice proposed [6, 8]. A image ehacemet algorithm is ideal oly whe it matches the uique image features; hece, whe referrig to a chemical elemet distributio image of the ureilite acquired from the Microprobe, the special features of these images eeded to be cosidered. After the cluster of iterested was segmeted, image cotet iside the cluster was kept. Iside this cluster, Our

4 ehacemet techique combies the Gaussia smooth filter together with a modified histogram equalizatio method to map the majority of the image itesities ito the maximum dyamic display rage available. We proposed to partially equalize the gray values Histogram equalizatio (HE) method. HE method is a statistical aalysis of image itesities. The histogram of a digital image couts the umber of pixels whose values are iside a set of rages called bis. If we use I to represet a image, a histogram is defied as a discrete fuctio ( low < I( x, high ),i= 1, H( = # pixels L The ( low, i i i, high i ) is the rage of the i-th bi, ad is the total umber of bis used. The relative frequecy histogram approximately defies the probability of occurrece at the i bi, as P( = H ( N,i= 1, L, where N is the total umber of pixels i a image takig ito cosideratios. A good ehacemet result will have the histogram stretched to the highest dyamic display rage. The histogram of pixels iside the selected cluster teds to gravitate to the lower itesity withi a arrow rage with outliers i both the high ad low ed i the low cotrast chemical elemet images that we aim to ehace. The outliers are classified dyamically as Outliers = { I lowest Bi(, if P( < ε, or if P( < ε } i= 1 i= highest with the lowest ad highest represet the ladmark values used to defie the outliers, which ca be set to about three times the iter quartile rage (3IQR) away from the media. I our case, we use the 0.1% data as outliers. After the outliers are classified, we set the I = lowest 1, if I = highest + 1, if lowest i= 1 P( < ε, ad I i= highest P( < ε, ad I Bi( Bi( I additio a Gaussia smooth filter is applied for preprocessig to make sure the outliers are properly classified ad a histogram of best fit ca be resulted. Our ew histogram image ehacemet steps are: 1. Determie the lowest 0.1% outliers i the histogram. 2. Determie the highest 0.1% outlier i the histogram. 3. Geerate ew histogram usig the itesity costraied betwee the lowest ad highest outliers defied. 4. Apply histogram equalizatio. (a) (b) (c) Figure 4: The Fe image segmet (a) ad the cluster idetified (b), followed by the ehaced clusters (c).

5 techique is applied; the cluster is marked red color. The resultig image i Fig. 4(c) has higher cotrast tha the image i Fig. 4(b). This is further show i the histogram i Fig. 5, where the histogram of the Fig. 4(b) is show i the Fig. 5 (b). The histogram of Fig. 4(c) is show i Fig. 5 (c) with Gaussia filter applied. Fig. 5 (c) that is correspoded to the partial HE method has show improved histogram with our partial HE method. Cluster segmetatio algorithm Fe Image Al Image (a) Mg Image cluster idetified Idetify the same cluster i the Fe ad Al images Localized ehacemet applied to both clusters Ehaced cluster image Fe Ehaced Al cluster image (b) Sythesized Color image output Figure 6. The diagram for geeratig the color sythesized cluster image. 5. SYNTHESIZED COLOR IMAGE I Fig. 4(a) is show the cluster of Fe image. I Fig. 4(b) is show the cluster of Fe image after the HE is directly applied; the cluster is marked i gree color. I Fig. 4(c), our partial HE ehacemet (c) Figure 5: The histogram of the origial Fe image cluster (a); the histogram of the cluster that histeq was directly applied (b); the histogram of our ehaced cluster (c). Huma eyes perceive color from three types of coes i their retia-red ( R), Gree (G) ad Blue (B), correspodig to the three basic colors RGB. A color image is represeted i (R, G, B) vectors. We ca differetiate just oticeable differece i terms of little chages i color distace, which makes the color image rich to iterpret. Here, we have composed a color image usig the three chemical elemet image clusters. The color images are geerated after the segmetatio ad ehacemet are obtaied. A color image is the composed of vector (R, G, B) for each image pixel. Here we desiged the color vector (R, G, B)=(Fe,

6 Mg, A. The whole processig is show is the diagram i Fig. 6. The sythesized color image is show i Fig. 7 (a), i compariso to the Mg gray itesity image. The yellow color show i the image likely idicated rich iro cotet regio. This distributio color allows us to further study the compoud distributio ad display the distributio of differet chemical compouds simultaeously i a color image. 6. CONCLUSIONS I this paper, we have proposed to localize the cluster of iterest i the meteorite images usig the Mg image as the referece. We further applied a image ehacemet to improve the visualizatio of the low cotrast images, ad evetually sythesized the three images of higher cotrast ito a color image after the segmetatio ad ehacemet. (a) 7. ACKNOWLEDGMENTS Our thaks to Dr. Steve Sigletary for his (b) help with the images provided. We also thak Miss Figure 7. (a) The color sythesized cluster image. (b) the Cassadra Hall ad Miss Siera Gozales. Both are gray itesity Mg image. studets participated i the CPSER program i Fayetteville State Uiversity, ad have cotributed to this research whe they are participated i the metored summer research project sposored by the Ceter for Promotig STEM Educatio ad Research. 8. REFERENCE [1] S. J. Sigletary ad T. L. Grove, "Early petrologic processes o the ureilite paret body," Meteoritics & Plaetary Sciece, vol. 38, pp , Ja [2] R. C. Gozalez ad R. E. Woods, Digital image processig, 2d ed. Upper Saddle River, N.J.: Pretice Hall, [3] J. Cay, "A Computatioal Approach to Edge- Detectio," Ieee Trasactios o Patter Aalysis ad Machie Itelligece, vol. 8, pp , Nov [4] Y. F. Bao ad H. Krim, "Smart oliear diffusio: A probabilistic approach," Ieee Trasactios o Patter Aalysis ad Machie Itelligece, vol. 26, pp , Ja [5] J. L. Lehr ad P. Capek, "Histogram equalizatio of CT images," Radiology, vol. 154, pp , Ja [6] J. S. Tag, et al., "Image ehacemet usig a cotrast measure i the compressed domai," Ieee Sigal Processig Letters, vol. 10, pp , Oct [7] J. A. Jiag, C. L Chuag, Y.L. Lu ad C.S. Fah, "Mathematical-morphology-based edge detectors for detectio of thi edges i lowcotrast regios", IET Image Processig, Vol 1 (3) pp , 2007 [8] C. Leug, K. Cha, H. Cha, W. Tsui, "A ew approach for image ehacemet applied to low-cotrast low-illumiatio IC ad documet images", Patter Recogitio Letters, Vol. 26 pp , [9] A. Yuille, T.A. Poggio, "Scalig theorems for zero crossigs." IEEE Tras. Patter Aalysis & Machie Itelligece, Vol. PAMI-8, o. 1, pp , Ja