Face Aging Modeling (Face Image Modeling and Representation) (Face Recognition) Part 3

Face Recognition Tests

The performance of the aging model is evaluated by comparing the face recognition accuracy of a state-of-the-art matcher before and after aging simulation. The probe set,is    constructed    by selecting one imagefor each subject i at age xi in each database,The   gallery    set    G    = is similarly constructed.

Table 10.2 Databases used in aging modeling [13]

Database		#subjects	#images	Average #images per subject
FG-NET		82	1002	12
MORPH	Albuml	625	l690	2.7
	Album2	4039	15204	3.8
BROWNS		4	132	33

Table 10.3 Probe and gallery data used in face recognition tests [13]

Database	Probe			Gallery
	#images	#subjects	Age group	#images	#subjects	Age group
FG-NET	82	82	(0, 5,…,30}	82	82	x* + (5, 10,…,30}
MORPH	612	612	(15, 20,…, 30}	612	612	x + (5,10,…,30}
BROWNS	4	4	(15, 20,…, 30}	100	100	x + (5, 10,…,30}

*x is the age of the probe group

A number of different probe and gallery age groups are also constructed from the three databases to demonstrate the model’s effectiveness in different periods of the aging process.

Aging simulation is performed in both aging and de-aging directions for each subject i in the probe and each subject j in the gallery as

Table 10.3 summarizes the probe and gallery data sets used in the face recognition test [13].

Let P, Pf and Pa denote the probe, the pose-corrected probe, and the age-adjusted probe set, respectively. Let G, Gf and Ga denote the gallery, the pose-corrected gallery, and age-adjusted gallery set, respectively. All age-adjusted images are generated (in a leave-one-person-out fashion for FG-NET) using the shape and texture pattern space. The face recognition test is performed on the following probe-gallery pairs: P-G, P-Gf, Pf-G, Pf -Gf, Pa-Gf and Pf -Ga. The identification rate for the probe-gallery pair P-G is the performance on original images without applying any aging model. The accuracy obtained by fusion of P-G, P-Gf, Pf -G and Pf -G f matchings is regarded as the performance after pose correction. The accuracy obtained by fusion of all the pairs P-G, P-Gf, Pf -G, Pf -G f, Pa-G f and Pf -Ga represents the performance after aging simulation. A simple score-sum based fusion is used in all the experiments.

Effects of Different Cropping Methods

The performance of the face recognition system is evaluated with different face cropping methods. An illustration of the cropping results obtained by different approaches is shown in Fig. 10.8. The first column shows the input face image and the second column shows the cropped face obtained using the 68 feature points provided in the FG-NET database without pose correction. The third column shows the cropped face obtained with the additional 13 points (total 81 feature points) for forehead inclusion without any pose correction. The last column shows the cropping obtained by the 81 feature points, with pose correction.

Fig. 10.8 Example images showing different face cropping methods: a original, b no-forehead and no pose correction, c no pose correction with forehead, d pose correction with forehead [13]

Figure 10.9(a) shows the face recognition performance on FG-NET using only shape modeling based on different face cropping methods and feature point detection methods. Face images with pose correction that include the forehead show the best performance. This result shows that the forehead does influence the face recognition performance, although it has been a common practice to remove the forehead in AAM based feature point detection and subsequent face modeling [3, 8, 26]. Therefore, the aging simulation is evaluated with the model that contains the forehead region with pose correction.

Note that, the performance difference between nonfrontal and frontal pose is as expected, and that the performance using automatically detected feature points is lower than that of manually labeled feature points. However, the performance with automatic feature point detection is still better than that of matching original images before applying the aging modeling.

Fig. 10.9 Cumulative Match Characteristic (CMC) curves with different methods of face cropping and shape & texture modeling [13]

Effects of Different Strategies in Employing Shape and Texture

Most of the existing face aging modeling techniques use either only shape or a combination of shape and texture [7, 8, 14, 20, 26]. Park et el. have tested the aging model with shape only, separate shape and texture, and combined shape and texture modeling. In the test of the combined scheme, shape and the texture are concatenated and a second stage of principle component analysis is applied to remove the possible correlation between shape and texture as in the AAM face modeling technique.

Figure 10.9(b) shows the face recognition performance of different approaches to shape and texture modeling. Consistent performance drop has been observed in face recognition performance when the texture is used together with the shape. The best performance is observed by combining shape modeling and shape + texture modeling using score level fusion. When simulating the texture, the aging simulated texture and the original texture have been blended with equal weights. Unlike shape, texture is a higher dimensional vector that can easily deviate from its original identity after the aging simulation. Even though performing aging simulation on texture produces more realistic face images, it can easily lose the original face-based identity information. The blending process with the original texture reduces the deviation and generates better recognition performance. In Fig. 10.9(b), shape + texture modeling represent separate modeling of shape and texture, shape + texture x 0.5 represents the same procedure but with the blending of the simulated texture with the original texture. The fusion of shape and shape + texture x 0.5 strategy is used for the following aging modeling experiments.

Effects of Different Filling Methods in Model Construction

Park et el. tried a few different methods of filling missing values in aging pattern space construction (see Sect. 10.3.1): linear, u-RBF, and RBF. The rank-one accuracies are obtained as 36.12%, 35.19%, and 36.35% in shape + texture x 0.5 modeling method for linear, u-RBF, and RBF methods, respectively. The linear interpolation method is used in the rest of the experiments for the following reasons: (i) performance difference is minor, (ii) linear interpolation is computationally efficient, and (iii) the calculation of RBF based mapping function can be ill-posed.

Figure 10.10 provides the Cumulative Match Characteristic (CMC) curves with original, pose-corrected and aging simulated images in FG-NET, MORPH and BROWNS, respectively. It can be seen that there are significant performance improvement after aging modeling and simulation in all the three databases. The amount of improvement due to aging simulation is more or less similar with those of other studies as shown in Table 10.1. However, Park et el. used FaceVACS, a state-of-the-art face matcher, which is known to be more robust against internal and external facial variations (e.g., pose, lighting, expression, etc.) than simple PCA based matchers. They argued that the performance gain using FaceVACS is more realistic than that of a PCA matcher reported in other studies. Further, unlike other studies, they have used the entire FG-NET and MORPH-Album1 in the experiments. Another unique attribute of their studies is that the model is built on FG-NET and then evaluated on independent databases MORPH and BROWNS.

Figure 10.11 presents the rank-one identification accuracy for each of the 42 different age pair groups of probe and gallery in the FG-NET database. The aging process can be separated as growth and development (age < 18) and adult aging process (age > 18). The face recognition performance is somewhat lower in the growth process where more changes occur in the facial appearance. However, the aging process provides performance improvements in both of the age groups, <18 and >18. The average recognition results for age groups <18 are improved from 17.3% to 24.8% and those for age groups >18 are improved from 38.5% to 54.2%.

Fig. 10.10 Cumulative Match Characteristic (CMC) curves [13]

 

 

 

Fig. 10.11 Rank-one identification accuracy for each probe and gallery age group: a before aging simulation, b after aging simulation, and c the amount of improvements after aging simulation [13]

Fig. 10.12 Example matching results before and after aging simulation for seven different subjects: a probe, b pose-corrected probe, c age-adjusted probe, d pose-corrected gallery, and e gallery. The ages in each (probe, gallery) pair are (0, 18), (0, 9), (4, 14), (3, 20), (30, 49), (0, 7) and (23, 31), respectively, from the top to the bottom row [13]

Matching results for seven subjects in FG-NET are demonstrated in Fig. 10.12. The face recognition fails without aging simulation for all these subjects but succeeds with aging simulations for the first five of the seven subjects. The aging simulation fails to provide correct matchings for the last two subjects, possibly due to poor texture quality (for the sixth subject) or large pose and illumination variation (for the seventh subject).

The aging model construction takes about 44 s. The aging model is constructed off-line, therefore its computation time is not a major concern. In the recognition stage, the entire process, including feature points detection, aging simulation, enrollment and matching takes about 12 s per probe image. Note that the gallery images are preprocessed off-line. All computation times are measured on a Pentium 4, 3.2 GHz, 3 G-Byte RAM machine.

Conclusions

A 3D facial aging model and simulation method for age-invariant face recognition has been described. It is shown that extension of shape modeling from 2D to 3D domain gives additional capability of compensating for pose and, potentially, lighting variations. Moreover, the use of 3D model appears to provide a more powerful modeling capability than 2D age modeling because the changes in human face configuration occur primarily in 3D domain. The aging model has been evaluated using a state-of-the-art commercial face recognition engine (FaceVACS). Face recognition performances have been improved on three different publicly available aging databases. It is shown that the method is capable of handling both growth and developmental adult face aging effects.

Exploring different (nonlinear) methods for building aging pattern space, given noisy 2D or 3D shape and texture data, with cross validation of the aging pattern space and aging simulation results in terms of face recognition performance can further improve simulated aging. Age estimation is crucial if a fully automatic age invariant face recognition system is needed.