An Automatic Weight-Based High Dynamic Range Imaging Syntheses with Multiple Different Exposed Low Dynamic Range Images

Transcription

1 49 An Automatc Weght-Based Hgh Dynamc Range Imagng Syntheses wth Multple Dfferent Exposed Low Dynamc Range Images 1 Jun-You Chen and 2 Chen-Chung Lu Abstract 1. Introducton Hgh dynamc range (HDR) magng offers the capture of fathful representatons of real world scenes to become a powerful technque n many areas, provdng the maxmum amount of detal data for radologsts to examne medcal MR mages. Now, HDR magng s wdely utlzed n survellance, remote sensng, and space research. Ths paper proposes a smple and effectve scheme to generate a sngle HDR mage combned by multple low dynamc range (LDR) mages of a scene wth dfferent exposures. The proposed algorthm frst dvdes each LDR mage nto non-overlappng blocks of the same sze, and then evaluate each block s average ntensty and the statstcal values of ntenstes of the whole LDR mages. The proposed approach gves weght for each block accordng to each block s average ntensty. The fnal output block s assgned as a weghted average of the nput blocks acqured at dfferent exposures. The expermental results show that the proposed approach has three major advantages: () the proposed algorthm s smple and effectve, () the proposed algorthm s tme savng, due to operatng completely on ntensty only, () there are no restrctons n the mage dynamc color ranges n the proposed algorthm. Keywords : Hgh dynamc range (HDR), low dynamc range (LDR), weght. *Correspondng Author: Jun-You Chen (E-mal: gn @yahoo.com.tw) 1 Tape College of Martme Technology, Tamsu, New Tape, Tawan 2 Department of Electronc Engneerng, Natonal Chn-Y Unversty of Technology The work was partally supported by the Natonal Chn-Y Unversty of Technology, Tawan, under research contract NCUT 14-R-CL In scence, the dynamc range s defned as a rato of the maxmum physcal measure to the mnmum physcal measure. In photography, the defnton of dynamc range depends on what the dynamc range refers to. The dynamc of scenes s the brghtness rato of the brghtest part to the darkest part, and the dynamc of dsplays s the rato of the maxmum lumnance to the mnmum lumnance emtted from the screen, and the dynamc of mages s the brghtness rato of the brghtest part to the darkest part of the mage [1, 2, 3]. An mage or a scene s defned to be hgh dynamc range (HDR) whle ts dynamc range extremely exceeds the dynamc range of the capture or dsplay devces. Real-world scenes are often HDR because they can always contan a very enormous range of lght ntenstes at the same tme. For nstance, n a sunny day we want to take a photograph that contans the nsde of a room and some exteror scenes vsble through wndows. In most cases, the exteror scenes wll be very brght due to the sun s drectly llumnatng, whle the llumnaton of the nsde of the room s far darker. In order to make features n the dark areas vsble, one may use hgher exposure to acqure an mage, whch dsplays approprately exposed detals n the room and renders the brght area saturated to lose all detals of the exteror scene. On the other hand, one may utlze lower exposure to properly expose the exteror scenes to make features n the exteror scenes vsble, but the nteror of the room wll be underexposed so to lose all detals of the room [4, 5, 6].

2 50 In fact, tradtonal low range dynamc (LRD) magng cannot completely dsplay all the detal features of a scene that are detected by human vsual system. On the other hand, wth the rapd progressons n computer graphng and dgtal magng technologes, people have been ncreasng nterests n hgh dynamc range (HDR) magng [7, 8]. HDR magng offers the capture of fathful representatons of real world scenes, so t becomes a powerful technque n many areas; n medcal magng, t provdes the maxmum amount of detal data for radologsts to examne medcal MR mages. In dgtal photography, t produces preferred pctoral mages, fathfully reproduces overall appearance of orgnal scenes and the contrast relatonshps between objects n the scene, and predcts the vsblty of specfc objects n a scene [9, 10]. Now, HDR magng s wdely utlzed n survellance, remote sensng, and space research. Many researches have shown that HDR magng has much better performances than that of LDR magng n these areas mentoned above. The man purpose of hgh dynamc range magng s to accurately represent the real world scenes wth a large range of brghtness from the brghtest sunlght to the darkest shadow. Now, there are many methods to obtan the so-called hgh dynamc range (HDR) mages:() created drectly wth Computer-generated mages;() obtanng from modern magng-hardware; advanced HDR-CCD dgtal cameras are combned wth two dfferent exposure CCDs so they can get the HDR that s four tmes of the tradtonal camera s HDR ; () generatng by combnng the nformaton from multple LDR mages taken at dfferent exposure settngs [11, 12, 13]. The thrd method s the most popular and most effectve, t detects saturated pxels n the mages and compensates these saturated pxels wth the pxels of the mages taken under other exposure condtons. People can easly set the optmal exposure condtons to mprove dynamc range through exposure tme control because the exposure tme can be controlled accurately. The most popular method to generate HDR mages s to sequentally take multple mages of the scene usng dfferent exposures. Hgh exposures provde useful nformaton n dark scene regons, yet low exposures offer useful nformaton n dark scene regons. Therefore, these dfferent exposed mages can be fused to get a sngle HDR mage. To acqure HDR mages from combnng multple LDR mages was frst reported by Mann & Pcard [14]. They examned the stuaton of each dfferent exposure mages of a scene to gve weght for each mage, and these weghted mages are merged to form a sngle HDR mage. Debevec and Malk [15] utlzed multple exposures to ncrease the dynamc range of mages. They gave hgher weght to nput pxels whose ntenstes are nearer to the mean of all the pxels of nput mages, and less weght to the nput pxels whose ntenstes are more far away from the mean of all the pxels of nput mages. The fnal output pxels are specfed as a weghted average of the nput pxels acqured at dfferent exposures. Chen and Mu [16] proposed an nteractve cut-and-paste scheme to ncrease mages dynamc ranges, where blocks of the resultant mage are manually selected from blocks of the nput mages. Addtonal works about HDR magng have been done by Mtsunaga & Nayar [17], Robertson et al.[18], and Robertson et al [19].We propose a smple and effectve scheme to generate a sngle HDR mage combned by multple LDR mages wth dfferent exposures of a scene; the method s an extenson to merge Chen et al method and Debevec et al method. Our scheme does not need to estmate the response functon of the mage capture devce. We frst dvde each LDR mages nto non overlappng blocks of the same sze, and then evaluate each block s average ntensty and the statstcal values of ntenstes of the whole LDR mages. Our algorthm gves hgher weght to blocks whose average ntenstes are nearer to the mean of all the correspondng blocks that are at the same place n the result mage, and less weght to blocks whose average ntenstes are nearer to the mean of all the correspondng blocks that are at the same place n the result mage. The fnal output block s assgned as a weghted average of the nput blocks acqured at dfferent exposures. The expermental

3 51 results show that the proposed algorthm s effectve and vald. The remander of ths paper s organzed as follows; Secton 2 presents the proposed algorthm. Secton 3 descrbes the emprcal results. Secton 4 concludes ths paper. 2. Proposed Object Extracton Algorthm An effcent hgh dynamc range magng scheme for color mages must be precse and tme-savng. In order to construct a superor hgh dynamc range magng algorthm for LDR color mages, several schemes are used n ths paper to acheve the goal. The overall process of the proposed hgh dynamc range magng scheme for color mages s shown n Fgure 1. The nput RGB color mages of a scene wth dfferent exposures are transformed nto HSI color space. The proposed scheme frst dvde each ntensty plane of HSI mages nto non overlappng blocks wth the same sze, and then evaluate each block s average ntensty and the statstcal values of ntenstes of the whole HSI mages. Our algorthm gve hgher weght to blocks whose average ntenstes are nearer to the mean of all the correspondng blocks that are at the same place n the result mage, and less weght to blocks whose average ntenstes are nearer to the mean of all the correspondng blocks that are at the same place n the result mage. The fnal output block s assgned as a weghted average of the nput blocks acqured at dfferent exposures. Some basc theory about human vsble system, the brghtness of a scene, and steps of the proposed algorthm are descrbed n detal n the followng subsectons. H,S SAWII DSRI Xs RGB 2 HSI I WCB CWB RIP Y DIUS EAIS Xs : dfferent-exposed low dynamc range mages of a scene Y : output the result hgh dynamc rang mage SAWII : statstcal analyss of whole mages ntenstes WCB : weght eaculaton of each sub blocks DIPB : dvdng Intensty plane nto unoverlapped sub blocks CWB : combnaton of weghted sub blocks EAIB : evaluatng the average ntensty of sub blocks DSRI : determnng the sutable range of ntensty RGB2HSI : color space transformaton from RGB doman to HIS doman RIP : resultant ntensty plane Fgure 1: The flow chart of the proposed hgh dynamc range magng scheme. 2.1 Human Vsual System and Scene Brghtness The human vsual system (HVS) can perceve about 4 orders of magntude of lght ntenstes at any one moment. The HVS can also be adjusted another 6 orders up and down through ts own adaptaton process. Ths adaptaton process of HVS does not work nstantaneously and may take several mnutes, for example, n the case of enterng a dark place from a brght envronment [8]. Even though the HVS can cover that huge range of lght ntenstes, t does not mean that people can see equally well at all ntensty levels, snce the HVS has two knds of photoreceptors, rods and cones, whch have dfferent senstvtes, respectvely. Rods can detect small lumnance dfferences n low-lght-envronments, but ther ablty for dstngushng colors s poor due to ther vsual sharpness. It s named scotopc vson. On the other hand, the three types of cones are conscentous for dstngushng a sharp vson of colors under well lghtness condtons. Ths s called photopc vson. Moreover, the overlappng regon of the scotopc and photopc ranges s called mesopc. The cones have an obstacle n dstngushng lumnance levels, whch have to be large enough to be detectble [10]. These

4 52 dfferences of detectble lght ntenstes were evaluated n psychophyscal studes. They have llustrated that over a large range of lght ntensty, the mnmum detectble lght dfference s almost constant around 1% [2]. The mnmum detectble lght dfference s called the vsble threshold or just-notceable dfference. Furthermore, researchers have shown that the HVS follows a logarthmc functon to response to lght ntenstes. The observed brghtness of a scene s referred as lumnance, measured n candela (cd) per square meters [2]. For example, the surface of the sun has an ntensty about cd/m², a moonless nght sky has a lumnance level about cd/m², and a daylght scene s close to 10 5 cd/m2 [4]. The dynamc range of brghtness (radance) values approxmates10 6 :1 for a typcal real-world scene, :1 for human eyes after adaptng, up to : 1 for camera sensor, and less than 10 3 : 1 for montor. Although the HVS has a dynamc range of brghtness approxmately twelve orders of magntude to dstngush huge range of lght ntenstes n the real world scene. But the current performance of modern dgtal mage capture and dsplay devces stll suffer from a lmted dynamc range. These devces are called low dynamc range (LDR) reproducton devces n ths paper. Dgtal cameras utlzed n computer vson usually offer 8 bts of brghtness nformaton for each color channel at each pxel. For each color channel, all radance ntenstes n the scene are mapped to one of 256 mage brghtness levels. Computer vson problems, such as the shape detecton of objects, the moton estmaton of objects, and the recognton of objects, are often under-controlled and thus essentally hard to be solved. The low dynamc mages produced by today s low dynamc range reproducton devces make dark areas darker and make brght areas brghter, losng more features to cause each of these problems more dffcult. 2.2 Color Model Transformaton from RGB to HSI For measurng or reproducng color, a number of three dmensonal color models are defned, among whch the most popularly used s RGB (Red, Green, and Blue) model [20]. The RGB model s a physcal system, and the mage n the RGB model s the most sutable for color mage representaton. However, t s not sutable for mage processng applcatons because ts R, G, and B components are hghly correlated. The dstance n the RGB color space does not stand for the perceptual dfference n a unform scale. In mage processng and analyss, these R, G, and B components are often transformed nto other color models. Modern technques for HDR magng are bascally developed on the RGB color space. Lumnance-chromnance color space representatons are frequently neglected. On the other hand, employng HDR magng technques n lumnance- chromnance space may be better for the followng reasons: () The ntensty channel of Lumnance- chromnance color space s the weghted average of the R, G, and B channels. It has a hgher sgnal-to-nose rato (SNR). () Lumnance-chromnance color space s a decorrelated color space so to offer better compressblty. Therefore, most part of mage compresson technques store mages n Lumnance-chromnance color space. Whle one utlzes the already-compressed multple-exposure LDR mages to make a HDR mage, t s more effcent and tme savng to generate the HDR mage, because the fnal HDR mage s more suted for compresson and dsplay. () HDR technques workng n RGB space always need post-composton whte balancng. The whte balancng should create perceptually convncng colors whch may not be the true colors. In ths paper, we address the HDR magng n HIS color space to mprove the performance of HDR magng [21].

5 53 The HSI (Hue, Saturaton, and Intensty) color model s the most representatve of the perceptual systems, whch s wdely used n mage processng. The advantages of the HSI model are: hue and saturaton have good correlaton wth the human percepton of colors and ts separablty of chromatc values from achromatc values. The HSI reduces the redundancy models n the RGB model, and ts components H (hue), I (ntensty), and S (saturaton) are gven usng some color transform from the RGB color space [22]. H 360 f B G f B G 1 R G R B 2 2 R G R BG B 1 (1) 1 cos 2 (2) S 3 1 [mn R, G, B] R G B (3) 1 I 3 R G B (4) In the HSI color model the saturaton corresponds to the relatve purty of a color. The hue stands for the domnant wavelength n mxed lght and ndcates a domnant color as perceved by the human eyes. The ntensty or perceved lghtness ndcates the brghtness of a color. 2.3 HDR Image Synthess Snce the hue and saturaton planes of dfferent exposure mages of a scene are the same, the proposed algorthm does not change the hue and saturaton planes to mantan ther orgnal colors. For estmatng the combnaton weght of each ntensty plane, the mean and standard devaton of ntensty of the set of dfferent exposure mages of a scene are evaluated by the followng equatons: W H x 1y 1 I ( x, y) /( W H), (5) N 1 / N, (6) N W H 1 x1 y1 H 2 ( I ( x, y) ) /( N W ), (7) where I ( x, y) s the ntensty of pxel p ( x, y) n the mage, W s the wdth and H s the heght of the nput mage, and N s the number of mages set of N dfferent exposure mages of a scene. The proposed algorthm dvdes each ntensty plane nto non overlappng blocks wth the same sze m n, evaluates each block s average ntensty, and then determnes each block s combnaton weght usng followng equatons. ( j, j m m 1 knn1 x jm ykn I ( x, y) /( m n), j 0,1,2,...,[ W / m], k 0,1,...,[ H / n], N ( j, exp( ( ( j, ) ) exp( ( ( j, ) ), where ( j, s the average ntensty of block ( j, of mage, ( j, s the combnaton weght of block ( j, of mage, and s the adjustment coeffcent wth value between 0.1 and 1. The ntensty of a block of the result HDR mage s the weghted sum of ntenstes of correspondng blocks located at the same place, and the ntensty plane of the resultant HDR mage s the composton of these result ntensty blocks. 3. Experment Result Ths secton presents expermental results under varous condtons to llustrate the utlty and effcency of the proposed scheme. Experments n ths paper are conducted on a computer wth a 2.8GHz Intel Pentum processor and 2 GB RAM runnng Matlab verson 7.6, and the nput RGB color mages are pxels. (8) (9)

6 54 Fgure 2 shows a hgh dynamc range magng example wth the proposed adaptve weghted sum (AWS) algorthm for a statc wndow scene. Row 1 shows fve nput LDR mages of a statc wndow scene taken at dfferent exposure tmes from brghtest to darkest,.e. 1/13, 1/25, 1/50, 1/100, and 1/200 seconds by usng a Nkon D5100 dgtal camera wth aperture f/8. Row 2 shows fve correspondng output HDR mages of these nput wndow scene mages generated at dfferent adjustment coeffcents from left to rght,.e. 0.2, 0.4, 0.6, 0.8, and 1. Row 2 shows that both ndoor and outdoor scenes are clearer whle the dfferent adjustment coeffcent s decreasng, so that more detal features can be detected. Row 3 shows four correspondng output HDR mages of the red-swatch mage of (c1) whle generated wth dfferent block sze from (c2) to (c5),.e. 3 3, 4 4, 6 6, and 8 8 pxels. Row 3 shows that the vsble qualty of the result HDR mage largely depends on the block sze, the edges of objects n the HDR mage s smoother whle the sze of blocks s decreasng. (a1) (a2) (a3) (a4) (a5) (b1) (b2) (b3) (b4) (b5) (c1) (c2) (c3) (c4) (c5) Fgure 2: A HDRI example of the proposed AWS algorthm for a statc wndow scene Table 1 shows the tme consumpton of the proposed AWS algorthm appled on varyng block szes and varyng number of nput mages. Table 1 shows that the tme consumpton s ncreasng whle the number of blocks s ncreasng. Moreover, the tme consumpton s ncreasng lttle whle the number of blocks s ncreasng. Table1: Tme consumpton (n seconds) of the proposed AWS algorthm vs. block szes and the number of nput mages Block sze mages

7 55 Fgure 3 shows another hgh dynamc range magng example wth the proposed adaptve weghted sum (AWS) algorthm for a statc scene; (a1) shows the nput over-exposure mage wth exposure value negatve one, (b1) shows the nput sutable exposure mage wth exposure value zero; (c1) shows the nput under-exposure mage wth exposure value postve one; (d1) shows the mage of hue plane of nput mage; (d2) shows the mage of the saturaton plane of nput mage; (a2) s the mage of ntensty plane of (a1); (b2) s the mage of ntensty plane of (b1); (c2) s the mage of ntensty plane of (c1); (a3) s the mage of weghted ntensty plane of (a1); (b3) s the mage of weghted ntensty plane of (b1); (c3) s the mage of weghted ntensty plane of (c1); (d3) s the mage of sum of weghted ntensty planes of nput mages; (d4) s the result HDR mage generated by the proposed AWS algorthm; (a4) s the mage of the absolute dfference between the result HDR mage and the nput over-exposure mage; (b4) s the mage of the absolute dfference between the result HDR mage and the nput sutable-exposure mage; (c4) s the mage of the absolute dfference between the result HDR mage and the nput under-exposure mage. Fgure 3 show that not only the vsble regons are preserved, but also the overexposed and underexposed areas n the orgnal LDR mages are modfed sutably n the result HDR mage. These facts show that the proposed algorthm can transform the source mage s colors to the target mage effectvely and accurately. (a1) (b1) (c1) (d1) (a2) (b2) (c2) (d2) (a3) (b3) (c3) (d3) (a4) (b4) (c4) (d4) Fgure 3: A hgh dynamc range magng example usng the proposed adaptve weghted sum (AWS) algorthm for a statc scene.

8 56 4. Conclusons HDR magng offers the capture of fathful representatons of real world scenes to become a powerful technque n many areas, so that t provdes the maxmum amount of detal data for radologsts to examne medcal MR mages. Now, HDR magng s wdely utlzed n survellance, remote sensng, and space research. Ths paper proposes a smple and effectve scheme to generate a sngle HDR mage combned by multple LDR mages of a scene wth dfferent exposures. The proposed algorthm frst dvdes each LDR mages nto non overlappng blocks of the same sze, and then evaluates each block s average ntensty and the statstcal values of ntenstes of the whole LDR mages. The proposed algorthm gves weght for each block accordng to each block s average ntensty. The fnal output block s assgned as a weghted average of the nput blocks acqured at dfferent exposures. The expermental results show that the proposed approach has three major advantages: () the proposed algorthm s smple and effectve, () the proposed algorthm s tme savng, due to operatng completely on ntensty only, () there are no restrctons n the mage dynamc color ranges n the proposed algorthm. In the future, we wll combne the proposed scheme color transform schemes to mprove the qualtes of generated HDR mages. References [1]. A. Srkantha, D. Sdbe, Ghost detecton and removal for hgh dynamc range mages: Recent advances, Sgnal Processng: Image Communcaton, Vol. 27, pp , [2]. M. A. Robertson, S. Borman, R. L. Stevenson, Estmaton-theoretc approach to dynamc range enhancement usng multple exposures, Journal of electronc magng, Vol. 12, No. 2, pp , [3]. S. Slk n, J. Lang, Hgh dynamc range mage deghostng by fast approxmate background modellng, Computers & Graphcs, Vol. 36, pp , [4]. L. Meylan, S. Süsstrunk, Hgh Dynamc Range Image Renderng wth a Retnex-Based Adaptve Flter, IEEE Transactons on mage processng, Vol. 15, No. 9, pp , [5]. B. Gu, W. L, J. Wong, M. Zhu, M. Wang, Gradent feld mult-exposure mages fuson for hgh dynamc range mage vsualzaton, J. Vs. Commun. Image R., Vol. 23, pp , [6]. B. Masa, S. Agustn, R. Flemng, O. Sorkne and D. Guterrez, Evaluaton of reverse tone mappng through varyng exposure condtons, ACM Transactons on graphcs, Vol. 28, No.5, [7]. A. Gonçalves, J. P. Moura, L. Magalhães, A. Chalmers, Perceptual mages of Conmbrga usng Hgh Dynamc Range, Journal of Archaeologcal Scence, Vol. 40, pp , [8]. G. Qu, J. Duana, G. D. Fnlaysonb, Learnng to dsplay hgh dynamc range mages, Pattern recognton, Vol. 40, pp , [9]. B. Gu, W. J. L, M. Y. Zhu, M. H. Wang, Local Edge-Preservng Multscale Decomposton for Hgh Dynamc Range Image Tone Mappng, IEEE Transactons on Image Processng, Vol. 22, No. 1, pp , [10]. S. Battato, A. Castorna, M. Mancuso, Hgh dynamc range magng for dgtal stll camera: an overvew, Journal of electronc magng, Vol. 12, No. 3, pp , [11]. T. Jnno, M. Okuda, Multple Exposure Fuson for Hgh Dynamc Range Image Acquston, IEEE Transactons on Image Processng, Vol. 21, No. 1, pp , [12]. F. Banterle, K. Debattsta, A. Artus, S. Pattanak, K. Myszkowsk, P. Ledda and A. Chalmers, Hgh dynamc range magng and low dynamc range expanson for generatng HDR content, Computer graphcs forum, Vol. 28, No. 8, pp , [13]. M. H. Km, J. Kautz, Characterzaton for hgh dynamc range magng, Computer graphcs forum (Proc. EUROGRAPHICS 2008), Vol. 27, No. 2, pp , [14]. S. Mann and R. Pcard, Beng undgtal wth dgtal cameras: Extendng dynamc range by combnng dfferently exposed pctures, IS&T s 48th annual conference, Cambrdge, MA, [15]. P. Debevec and J. Malk, Recoverng hgh dynamc range radance maps from photographs, SIGGRAPH 1997: Proceedngs of the 24th annual conference on Computer graphcs and nteractve technques, pp , 1997.

9 57 [16]. Z. Chen and G. Mu, Hgh-dynamc-range mage acquston and dsplay by mult-ntensty magery, J. Imagng Sc. Technol. Vol. 39, No. 6, pp , [17]. T. Mtsunaga and S. Nayar, Radometrc self calbraton, In Proceedngs of IEEE CVPR, pp , [18]. M. A. Robertson, S. Borman, and R. L. Stevenson, Dynamc range mprovement through multple exposures, Int. Conf. Image Process.3, , [19]. M. Robertson, S. Borman, and R. Stevenson, Dynamc range mprovements through multple exposures, Proceedngs of Internatonal Conference on Image Processng (ICIP) 1999, pp , [20]. C. Clausen, H. Wechsler, Color Image Compresson Usng PCA and Back propagaton Learnng, Pattern Recognton, vol.33,pp , [21]. O. Prnen, A. Fo, A. Gotchev, Color hgh dynamc range (HDR) magng: The lumnance-chromnance approach, Internatonal journal of magng systems and technology, Vol. 17, No. 3, pp , [22]. M. Sezgn, B. Sankur, Survey over mage thresholdng technques and quanttatve performance evaluaton, Journal of Electronc Imagng, vol. 13, no. 1, pp , Chen-Chung Lu receved the B.S. degree n appled mathematcs from Natonal Chung-Hsng Unversty, Tawan, n 1976, and M.S. degree n physcal oceanography from the Natonal Tawan Unversty, Tawan, n He spent two years (82/83) at the Natonal Tawan Natural Scence Museum as an assstant researcher. At present he s a Professor and the head n the Department of Electronc Engneerng at Natonal Chn-Y Unversty of Technology, Tawan. Hs research nterests nclude computer graphcs, pattern recognton, mage analyss, and dgtal sgnal processng. Jun-You Chen receved the B.S. degree n electronc engneerng from Natonal Chn-Y Unversty of Technology, Tawan, n 2010, and M.S. degree n electronc engneerng from Natonal Chn-Y Unversty of Technology, Tawan, n At present he s a lecture at Tape College of Martme Technology, New Tape, Tawan. Hs research nterests nclude computer graphcs, 3D model watermarkng, pattern recognton, mage analyss, and embedded system.