Comparison Of Synthetic Face Aging To Age Progression By Forensic .

progressed to different ages uses AAMs and is similar tothat in earlier work in growth and development [14] andalso in automatic adult age progression [17]. An AAM isgenerated based on performing principal componentsanalysis (PCA) on parameters of a shape model and atexture model that were each created using PCA onindividual shape (landmark location) and texturecoordinates from the image set [13]. AAM parametersrepresent faces using the distance from the averageshaped and textured face and may be used to classifyfaces or synthesize new faces within face space [11, 12].Figure 1: Age-progression by notable forensic artist of theprimary author (currently age 34) at ages 40, 50, 60, and 70.3.2. Synthetic TechniqueFamily images were compiled for the primary author ofthis work in a similar manner as would be conducted toprovide a forensic sketch artist. These images were usedto build a “family face space” using AAM techniques.Images were included from the author in his twenties andearly thirties as well as images throughout all availableages for his parents and both sets of grandparents. Asample of family images is shown in Figure 2.Figure 3: Facial landmarks used for AAM.3.2.2Simulation of Aging in ImagesSimulating aging of the faces in images was performedusing two steps. The first is to estimate the age of a face.The second involves shifting the AAM parameters in thedirection of a theoretical “aging” axis.Estimating the age requires a solution to the equationageest W1b W2b2 offset,(1)where b and b2 are the fifty-five AAM parameters and theparameters squared for the given image, W1 and W2 are aset of weights to be found to shift the model parameters,and offset is a constant to place values in an appropriaterange. Representing this equation as an optimizationproblem, a genetic algorithm (GA) was used to find anappropriate solution [16, 17]. The determined agingfunction could estimate ages very well in the younger andmiddle ages where there were more images used fortraining. After increasing the influence of the olderimages used for training, though, by adding them to theAAM building multiple times, improvements were madeto where the GA was quite accurate on all the images.Table 1 shows results for one of the images of the authorat age thirty-four.The second step of simulating aging may berepresented by the equationFigure 2: Sample of family images used.Ninety-nine images were used. Since there were fewerimages at older ages, however, several of the images offamily members at older ages were repeated in buildingthe AAM to prevent skewing toward younger images. Intotal, one-hundred and sixty-one landmarks on the face asshown in Figure 3 were used to create the active shapemodels and have texture samples warped to create theactive texture models. The landmarks used in this workare a superset of common anthropometric points and wereexpanded from a smaller, earlier set in an attempt to bettermodel regions that were likely to change during ageprogression [17]. Together these models were reduced tofifty-five AAM parameters using the typical PCA methodof combining shape and texture coordinates.b2 f(agenew, b1),(2)where b1 is the vector of parameters for a given facialimage, agenew is the desired age to which to transform the3.2.1Aging Using Active-Appearance ModelsThe method used to create the synthesized images249

Figure 4: Average family face age-progression sequence.face image, and b2 is the vector of parameters generatedfor the new age. The approach taken in this work is thecreation of a lookup table for a “generalized” agingmodel, in this case based solely on the aging of familymembers. One hundred thousand random but plausibleimages were generated using AAM parameters. Theseimages were averaged for each given age and used topopulate the table with the average face parameters for agiven age. This table represents the generalized agingmodel and helps illustrate some of the trends that the GAlearned in training on the range of ages of faces. Sampleimages from this are shown in Figure 4. Although imagestend to be smoothed somewhat by the AAM technique,there is noticeable texture darkening in the areas of theface discussed earlier concerning aspects of the agingface. Shape also changes slightly, but this is better seenin a succession of images or animation based on these.To synthesize the image of an individual at a specifiedage, the original image AAM parameters are transformed.The age of the individual in the image is estimated usingthe method described above, and the difference of ageparameters at the desired age and estimated age is takenfrom the lookup table. This difference vector is added tothe original parameters to generate parameters for animage of that individual at the target age. The resultingface image is synthesized using those parameters.The “aged” images produced a definite appearance ofolder faces. Changes in the areas noted in theanthropological literature, such as increased nasiolabialand lateral orbital lines as well as small shifts in faceshape are present. Slight shifts in the location ofparticular features are more easily noted when images areviewed in succession, and texture changes are noted aswell, although fifty-five AAM parameters still may notrepresent sufficient texture information for areas such asthe forehead to render creases properly. Overall, theAAM model technique does have a smoothing effect onimages.Actual Ageoriginal 34progressed 40progressed 50progressed 60progressed 70GA AgeEstimate33.239.748.957.965.4Table 1: Sample genetic algorithm estimates of original andage-progressed image.4. Results and Discussion4.1. Synthesized Age ProgressionA variety of images were age-progressed. Images ofthe author were “aged” and images of his father andgrandfathers were “de-aged.” Images were used to makequalitative comparisons and also used in a survey ofthirty-seven individuals. This survey was used to makesome informal, quantitative judgments of perception ofthe ages of individuals generated by the AAM methodand by the sketch artist.Figure 6: Author’s grandfather age-progressed in reversefrom age 55 to age 25.De-aged images demonstrate slight changes in location offeatures as expected and some reduction of texturedarkness for lines and wrinkles. Further resolution inboth aging and de-aging constructions, however, wouldlikely improve visually age-progressed images producedwith this technique.The synthesized images were compared with thosegenerated by the sketch artist. Most informal responsesindicated that the synthesized images looked more like theauthor, but that the sketched images demonstrated moreaging effects. Figure 7 shows a variety of similar imagesfor direct comparison at ages 40, 50, 60, and 70 from leftto right.Figure 5: Author age-progressed from ages 23 and 34respectively to ages 40, 50, 60, and 70 from left to right.250

work has shown that texture information for distinctwrinkles tends to produce a perception of older age byindividuals agedImages11.8yearsTable 2: Average error of age estimate by sample group.The GA performed better on age estimation on bothoriginal images and aged and de-aged images than thegroup average. These average error estimates are shownin Table 3. It has likely learned a different set ofindicators or weighting than those used by humanperception.Figure 7: Comparing various age-progressed images tosketches at ages 40, 50, 60, and 70.Figure 8 includes a variety of synthetic images of theauthor at age seventy. Informally, when viewing theseimages next to images of the same face at younger ages,changes were difficult to perceive and a younger agetended to be estimated, but when viewed separately or inanimated succession, changes were noticed and ages wereusually perceived to be older.OriginalReconstructed Aged rsyearsyearsTable 3: Average error of age by genetic algorithm.Finally, composite images were made wheresynthesized faces were placed back into an image with earand hair. Hair was receded similarly to family members.These images are included for comparison in Figure 9.Figure 8: Various synthetic renderings at age 70.Table 2 includes results from the survey conductedusing thirty-seven individuals. Previous work has shownaccurate results in age perception by individuals, but it isaffected greatly by texture and other features [10]. Thegroup was shown forty images for approximately tenseconds each and asked to intuitively estimate age withoutanalysis.The images included a variety of real,synthesized, and sketched images. The group average ofage estimation was shown to be quite accurate, with anaverage age estimate error of only 2.6 years on original,unaltered face images. The group had a larger error onthe sketches, however, with a group average error of 10.2years. Several images were simply reconstructed usingun-shifted AAM parameters. The group was successfulwith these as well, with a slightly larger error than that oforiginal images. The synthetic age-progressed imageshad the worst group average error of 19.5 years, tendingto estimate the images much younger than the intendedage progression, although “aged” images were typicallyperceived as older individuals. This is likely due to tworeasons: the first is the reduced texture detail due to themodel smoothing using only fifty-five parameters torepresent a face, the second is that the author was presentat the beginning of the study which may haveunconsciously affected the estimation of ages of “aged”images of the author. The de-aged image average errorwas better, closer to that of the forensic sketches. Thispossibly supports the previous theory. Also, previousFigure 9: Composite using image age-progressed to 70,sketch at 70, and composite using average family face at 70.4.2. Discussion of Images and SurveyThe increase in the number of landmarks used and thechange of location of landmarks to better model faceaging, as well as an increase in the number of AAMparameters, did notably improve the visual quality of theimages versus earlier work [14]. Initial work with thefamily images favored younger looking images due tofewer older images used in building the AAM. Therepetition of older images to increase their weighting inthe model building resulted in more visually apparent ageprogression.The images generated do qualitatively demonstrate anincrease in age, although it is difficult to quantitativelyexpress this. Texture information does still seem to beoverly smooth, as is evidenced by the error in ageestimation by the group.The group survey does seem to indicate that people aregood at estimating ages as previous work has suggested[10]. Mild shape change, though, does seem to have lessinfluence on human perception than texture. Improved251

texture representation will likely improve individuals’perceptions of synthetically aged images as the intendedtarget age. The GA constructed for the aging techniquealso seems to perform well. It is more accurate on theaged images, perhaps reflecting that it is picking up moreon changes of shape and requires less texture information.Psychology: Human Perception and Performance,vol. 1, no. 4, pp. 374-382, 1975.[9] A.J. O’Toole, T. Vetter, H. Volz, and E. Salter,“Three-Dimensional Caricatures of Human Heads:Distinctiveness and the Perception of Age,”Perception, vol. 26, pp. 719-732, 1997.5. Conclusion[10] D.M. Burt and D.I. Perrett, “Perception of Age inAdult Caucasian Male Faces: Computer GraphicsManipulation of Shape and Color Information,” Proc.Royal Soc. London, vol. 259, pp. 137-143, 1995.This method generates images that seem comparable tothat of a forensic sketch artist in some epresentation should yield better results. Future workwill seek to increase texture resolution and also betterquantify the comparison of methods. Other possibilitiesinclude incorporating nonlinear shifts of parameters andthe use of individualized aging functions and agingfunctions and methods that include statistically significantguidance of idiosyncratic behavior and other complicatingfactors such as weight gain or loss that may improve theaccuracy of synthetically age progressed images.[11] T.F. Cootes and C.J. Taylor, “Statistical models ofappearance for medical image analysis and computervision,” Proceedings of SPIE Medical Imaging, 2001.[12] G.J. Edwards, C.J. Taylor, and T.F. Cootes,“Interpreting Face Images Using Active AppearanceModels,” Proc. Fifth European Conf. ComputerVision, pp. 581-595, 1998.[13] A. Lanitis, C.J. Taylor, and T.F. Cootes, “AutomaticIdentification and Coding of Human Faces UsingFlexible Models,” IEEE Transactions on PatternAnalysis and Machine Intelligence, vol. 19, no. 7,July 1997.References[1] K. T. Taylor, “Forensic Art and Illustration,” BocaRaton: CRC Press, 2001.[2] M. Y. Iscan, Global forensic anthropology in the 21stcentury. Forensic Science International, Vol. 117,Issue 1-2, pp. 1-6, 2001.[14] A. Lanitis, C.J. Taylor, and T.F. Cootes, “TowardAutomatic Simulation of Aging Effects on FaceImages,” IEEE Transactions on Pattern Analysisand Machine Intelligence, vol. 24, no. 4, April 2002.[3] M.A. Taister, S.D. Holliday, and H. Borman,“Comments on Facial Aging in Law EnforcementInvestigation,” Forensic Science Communications,2000.[15] T.F. Cootes, C.J. Taylor, D.H. Cooper, and J.Graham, “Active Shape Models – Their Training andApplication,” Computer Vision Graphics and ImageUnderstanding, vol. 61, no. 1, pp. 38-59, 1995.[4] R. G. Behrents, “An Atlas of Growth in the AgingCraniofacial Skeleton,” Craniofacial Growth Series,Ann Arbor, Michigan, 1985.[16] D.E. Goldderg, Genetic Algorithms in SearchOptimization and Machine Learning. AddisonWesley, 1989.[5] M. S. Zimbler, M. Kokosa, J. R. Thomas, “Anatomyand Pathophysiology of Facial Aging,” Facial PlasticSurgery Clinics of North America, vol. 9, pp. 179187, 2001.[17] E. Patterson, K. Ricanek, M. Albert, and E. Boone,“Automatic Representation of Adult Aging in FacialImages,” IASTED International Conference onVisualization, Imaging, and Image Processing, Spain,August 2006.[6] A. Lanitis, C. J. Taylor, and T.F. Cootes, “SimulatingAging on Face Images,” Proc. Of the 2nd InternationalConference on Audio and Video-based BiometricPerson Authentication, 1999.[18] IAI Forensic Certifications. Retrieved May 2007from http://www.theiai.org/certifications/[7] J. B. Pittenger, R. E. Shaw, and L. S. Mark,“Perceptual Information for the Age Level of Facesas a Higher Order Invariant of Growth,” Journal ofExperimental Psychology: Human Perception andPerformance, vol. 5, no. 3, pp. 478-493, 1979.[19] B. Vaughan, “Man of the Month: Frank Bender,”Esquire, April 2004.[20] K. Ramsland, “Frank Bender: The Art of Crime,”Crime Library. Retrieved May 2007 fromhttp://www.crimelibrary.com/criminal mind/forensics/bender/6.html[8] J. B. Pittenger and R.E. Shaw, “Aging Faces asViscal-Elastical Events: Implications for a Theory ofNonrigid Shape Perception,” J. Experimental252

provide a forensic sketch artist. These images were used to build a "family face space" using AAM techniques. Images were included from the author in his twenties and early thirties as well as images throughout all available ages for his parents and both sets of grandparents. A sample of family images is shown in Figure 2.