Iranian Face Database With Age, Pose and Expression

Transcription

1 Iranian Face Database With Age, Pose and Expression Azam Bastanfard, Melika Abbasian Nik, Mohammad Mahdi Dehshibi Islamic Azad University, Karaj Branch, Computer Engineering Department, Daneshgah St, Rajaee Shahr, Karaj, Iran {melika_nik, Abstract This paper introduces a database of over 3,600 color images which we collected from 616 different human faces. It includes facial images of people between ages Face images of persons with different ages is needed to generate a reliable age classification algorithm. Four different poses, two different expressions, and one image with glasses were taken if the subjects were interested. Information such as career, kind of skin, fingerprint, and cosmetic points (surgical points, fracture or laceration on face) is taken from participants. The background was a white frame and the only illumination was the daylight. We named this database the Iranian Face Database (IFDB). To evaluate the database the experimental result of a new feature detection algorithm is reported. Index Terms face image database, age classification, feature detection algorithm. I. INTRODUCTION Human face is the most common and useful key to a person's identity. As humans, we are able to categorize a person s age group from a person s face image and are often able to be quite precise in this estimation [1]. In recent years, face recognition and related works have received substantial attention from researchers in biometrics, pattern recognition, and computer vision communities [2], [3], [4], [5]. These common interests among researchers working in diverse fields motivated us to collect a database of facial images from people of different ages. The database is intended for distribution to researchers. 1 There are many publically available databases for face recognition and facial expression analysis. Beside above applications, Iranian Face Database (IFDB) can be used for age classification, facial surgery, race detection (beside other databases), studying influence of career, kind of skin on aging, and other similar researches. The remaining part of the paper is organized as follows: Section II gives details of the existing face image databases and the need for a new database is considered in section III. Section IV describes the Iranian Face Image Database in detail and Section V goes on to evaluate the database by experimental result of a new feature detection algorithm. The concluding remarks are present in Section VI. 1 II. EXISTING FACE IMAGE DATABASES Many face databases are recorded under a variety of conditions and with various applications in mind. Along with the development of face recognition and expression algorithms, comparatively a large number of face databases have been collected; such as AR [6], Yale [7], MIT [8], JAFFE [9], and many other databases [10]. Here FERET [11], CAS-PEAL [12], CMU-PIE [13], [14] and FG-NET [15] databases are reviewed. A. FERET Database The Facial Recognition Technology (FERET) database was collected at George Mason University and the US Army Research Laboratory facilities as part of the FERET program [11]. In FERET database images of 1199 subject exist in 9-20 different poses, 2 facial expressions and 2 different illuminations in 2 different times. There are 14,051 images in pixels in size. Images were collected at the following poses: right and left profile, right and left quarter profile, and right and left half profile. In these categories images were recorded for 508 to 980 subjects. In addition, five irregularly spaced views were collected for 264 to 429 subjects. In a new release of the FERET database, NIST (National Institute of Standards and Technology) is making color images of most of the original gray-scale images available in higher resolution ( ) [10]. B. CAS-PEAL Database The CAS-PEAL (pose, expression, accessory, and lighting) Chinese face database was collected at the Chinese Academy of Sciences (CAS). It contains images of 1040 subjects (595 men, 445 women) [12]. Nine cameras located in a semicircle around the subject were used. Subjects were asked to look up and down for additional recordings resulting in 27 pose images. To record faces under different lighting conditions, constant ambient illumination together with 15 manually operated fluorescent lamps were used. For expression, subjects were asked to smile, frown, look surprised, close their eyes, and open their mouth. A small number of subjects were recorded wearing three types of glasses and three types of hats. To capture the effect of the camera auto white-balance, subjects were also recorded with five uniformly colored backgrounds (blue, white, black, red, and yellow). In addition, images were obtained /07/$ IEEE 50

2 at two distances (1.2 and 1.4 meters). Finally a small number of subjects returned 6 months later for additional recordings. Of the 99,594 images in the database, 30,900 images are available in the current release. Images are grayscale and pixels in size [10]. C. CMU-PIE (Pose, Illumination, and Expression) Database The CMU (Carnegie Mellon University) PIE database systematically samples a large number of pose and illumination conditions along with a variety of facial expressions. The PIE database contains 41,368 images obtained from 68 individuals [13]. The subjects were imaged in the CMU 3D Room [14] using a set of 13 synchronized high-quality color cameras and 21 flashes. The RGB color images are in size. In addition, by combining two illumination settings, a total of 43 illumination conditions were recorded. The subjects were asked to display a neutral face, to smile, and to close their eyes in order to simulate a blink. 60 frames of subjects talking were recorded using three cameras (frontal, threequarter, and profile views). D. FG-NET Aging Database The FG-NET Aging Database was generated as part of the European Union project FG-NET (Face and Gesture Recognition Research Network).This database contains 1002 scanned face images showing 82 subjects at different ages. Images have various resolutions; approximately pixels. The database was developed in an attempt to assist researchers who investigate the effects of aging on facial appearance [15]. III. NEED FOR A NEW DATABASE In order to build, train, and reliably test age classification algorithms, databases with controlled variations of factors such as age, face pose, facial expression, occlusion facial hair, and illumination is needed. In spite of various databases, there is not an appropriate one for age classification. Most current databases don t have images of people in different ages, or they do not mention their ages. FG-NET database contains scanned images of subjects with mentioning their ages; but different lighting conditions, background, poses, and expressions. So, it was concluded to provide a database with conditions of age classification project. Age, enough resolution for wrinkle analysis, and frontal poses are the basic needs for this field. IV. DESCRIPTION OF IRANIAN FACE DATABASE The Iranian Face Database, the first image database in middle-east, contains color facial imagery of a large number of Iranian subjects. It was collected in January and February 2007 in the Department Of Engineering, Islamic Azad University of Karaj. IFDB is a large database that can support studies of the age classification systems. It contains over 3,600 color images corresponding to 616 people's faces (487 men, 129 women), just male images are available. No restrictions on wear (clothes, glasses, etc.), make-up, hair style, facial hair were imposed to participants. Ground-truth information, including ID, age, kind of pose or expression and if the subject has glasses is provided. Experimental subjects were photographed with a fine-resolution color digital camera in daylight. The subjects were seated on a stool and instructed to maintain a constant head position (although slight movements were unavoidable). To construct the database, subjects with different ages were needed. We couldn t ask a large number of people to come to a defined studio; so it was decided to start photography from kindergartens, schools and universities to gather images of subjects between 2 and 30 years old. Then offices, departments, administrations, and parks are tried to find young, middle-aged and senior subjects. The images are in pixels resolution, 24 bit depth, about 40 Kbytes size and JPEG format. A. Illumination and Background Enough luminosity for wrinkle processing and facial features without shadows is needed (in age classification wrinkle detection and analysis is important for the distinguishing of seniors from those in the younger categories [1]). Subjects were photographed without any projectors or flashes in daylight. Because the recording equipment had to be reassembled for each session, some variations between recording sessions are present. Note that all images of each person were taken on the same session. The background is a m 2 white frame. It was designed so that it can be folded at the mid. B. Pose and Expression Subjects were imaged in natural frontal pose (without expressions and glasses) twice. They may have beards, moustaches, or scarf. If the person uses glasses, an image with glasses was taken too. Figure 1 shows an example of a frontal face image without and with glasses. Figure 1 An example of a frontal face image without and with glasses; Images that were taken by asking the subject to smile and frown ; Images of the subject with head up, head down, left profile and right profile. 51

3 If the subject was interested, images with poses and expression were taken too. To facilitate the collection process, subjects were asked to perform the desired actions. To capture images with pose variation, the subject was asked to look upwards, look downwards, and rotate head and body for right and left profile as shown in figure 1. There is no restriction for head angle. The human face is able to display an astonishing variety of expressions. Happiness, sadness, fear, disgust, surprise and anger, from either single images or image sequences are expressions what Ekman and Friesen called the six basic emotions [16]. In addition to the natural expression, cooperative subjects were asked to smile and frown as shown in figure 1. Totally, images with two natural frontal, 1 natural frontal with glasses, four poses and two expressions were exist in this database; however subjects may participate in some cases. In IFDB each subject has a file consists of his images and a text file including, kind of skin, career, and cosmetic points (surgical points, fracture or laceration on face). V. EVALUATION WITH A NEW FACIAL FEATURE DETECTION ALGORITHM To evaluate this database, a new detection algorithm is designed to find facial features. These features are shown in Figure 2. The localization of the facial features is performed in stages. At each stage, a particular facial feature parameter is found. Region of the face in the image is initialized manually, with an allowance for a large margin of error. Eyes are often analyzed for feature extraction algorithms. So at first eyes are found in 3 steps, then according to the distance of two eyes and the ratios between features [17], other components of face are extracted. A. Eye Finding Stage It consists of an eye region finding stage, eye region correction stage, and iris finding stage. First, face must be cropped from the image. As shown in Figure 2, it is cropped from below the line of hair growth, including sides of the face and it continues beneath the chin. Usually, eyes located in the middle height and above the face images. If W denotes width of cropped face image and H denotes height, the eyes searching region is located, respectively at W / 10 and 9W / 10 of the image from the left and 3H / 20 and 12H / 25 from the top of the image [18]. Searching region divided into two areas for left and right eyes. The MIPF v (Vertical Mean Integral Projection Function) and HIPF v (Vertical Hybrid Integral Projection Function) [18] are performed on the eyes searching areas. Top of the new region is calculated as follow: 1. Find the major valley on MIPF v ; 2. Find the major peaks at right or left point of major valley; 3. If maximum point in the HIPF v is close to the right major peak or the major valley of MIPF v, then: 3.1. If maximum point of MIPF v is close to the major left peak then choose the major valley as top of the new region; 3.2. Else choose the major left peak; 4. Else choose the major right peak; This point (top of the new region) is located at the bottom position of the eyebrow, as shown in Figure 3. Then spaces above this point will be removed from eye searching region. The PCA (Principal Component Analyses) [20], [21] is applied on the new eyes searching areas. The major peak is located at the iris s row position of the region, as shown in Figure 4. Then, PCA is applied on 90 rotated region and HIPF h (Horizontal Hybrid Integral Projection Function) [19] is applied on the eyes searching area. The major peak is located at the iris s column position of the region, as shown in Figure 5 and 6. If the HIPF h s point is close to the PCA s point, then PCA s point is selected as iris s column position; else HIPF h s point will be selected as iris s column position. So iris s column and row is found in this way. Figure 3 Vertical Mean Integral Projection on left eye region and the bottom position of the eyebrow. Figure 2 A cropped image and face template. 52

4 Figure 7 Distance between two irises. Figure 4 PCA on left eye region and the iris s position. Figure 5 Horizontal Hybrid Integral Projection on the left eye region and iris column coordinate. According to the facial model in [17], the distance from eyes to nose is around 0.6*D eyes and the left and right margins of the nose are located about the center of their corresponding eyes, respectively. The nose s searching region could be defined as illustrated in Figure 8. The top boundary is located at the 0.45*D eyes lower than two eyes and the bottom boundary is positioned at 0.4*D eyes downward from the top boundary. Similarly, first extra space from this region is removed by applying the vertical mean integral projection function (MIPF V ) on the nose s searching region, as shown in Figure 9. At the positions of the upper and bottom of nostrils two obvious valleys are revealed. Two valleys are beside highest peak. The one is close to upper nostril bound and the other is that of the bottom nostril bound and major peak in the PCA, as shown in Figure 10, is the nostril position. Mean of two nostril points is considered as nose point as shown in Figure 2. Figure 8 Nose s searching region. Figure 6 PCA on the left eye region (90 rotated) and iris s column coordinate. In 76.23% of images both eyes are found correctly and in 8.11% of images just one eye is found correctly. Other stages of finding feature are applied to the images with correct found eyes. The average distance between two eyes is about 118 pixels. B. Virtual Top of Head Finding Stage The top of the skull is difficult to estimate when the person has hair. Hence, semi automatic method for finding face region in image is making it easy. This point is positioned between two eyes and on the top boundary of cropped image as shown in Figure 2. C. Nose Finding Stage The nose in an image could be located in a similar way as eyes. Let D eyes be the distance between the irises coordinates of the eyes, as depicted in Figure 7. Figure 9 MIPF V' Upper and bottom bound of the nostril. Figure 10 PCA on Nostril position. 53

5 D. Lip Finding Stage The distance from eyes to mouth is around D eyes [17] and the left and right margins of the lip are located about the centers of their corresponding eyes, respectively. The lips searching region could be defined as illustrated in Figure 11. The top boundary is located at the 0.85*D eyes lower than two eyes and the bottom boundary is positioned at 0.45*D eyes downward from the top boundary. The vertical general projection function (GPF V ) [19] is performed on the lips searching area. The major valley is formed at the lip central line position of the region, as shown in Figure 11. The major peak in the PCA before lip central line position, as shown in Figure 12, is the lip upper bound position. E. Chin Finding Stage The distance from eyes to chin is around 1.5 *D eyes [17] and the left and right margins of the chin are located about the centers of their corresponding eyes, respectively. The chin s searching region could be defined as illustrated in Figure 13. The top boundary is located at the 1.4*D eyes lower than two eyes and the bottom boundary is at end of cropped image. PCA is performed on the chins searching area. The major peak after main valley, as shown in Figure 13, is formed at the chin line position of the region. Figure 13 - Chin searching area and PCA on the area. VI. CONCLUDING REMARKS In this paper, the development of a color image database for age classification is described. The database has been developed with the intention of improving researches about age classification and age estimation. It contains faces images with a large variety in age (see Figure 14) and images that have a good resolution and illumination for wrinkle analysis. Also images with different pose and expression (see table II) were taken for facial expression and other related researches. Totally images of 482 subjects are available for researchers now. To evaluate this database, a new feature detection algorithm is designed. This algorithm is applied on 616 facial images of IFDB, and the experimental results are pleasing. In table II result of this algorithm is shown. Figure 11 The lips searching region and GPF V on the searching region and lip center line position. Figure 14 Age histogram. Figure 12 PCA on the lips searching region and upper lip bound position. TABLE I NUMBERS OF EXPRESSIONS, POSES, AND OCCLUSION IMAGES IN IFDB State Numbers With Glasses 81 Right Profile 532 Left Profile 533 Head Down 271 Head Up 269 Anger 147 Smile

6 TABLE II RESULTS OF APPLYING OUR ALGORITHM ON IFDB component Percent of success two irises one eye 8.11 nose lip chin REFERENCES [1] Y. H. Kwon and N. da Vitoria Lobo, Age classification from facial images, Computer Vision and Image Understanding Journal, Vol. 74, No. 1, pp. 1-21, April [2] A. Bastanfard, H. Takahashi and M. Nakajima, Facial age declination based on medical anthropometry, Proc. NICOGRAPH International, pp.67-72, [3] R. Chellappa, C.L. Wilson, and S. Sirohey, Human and machine recognition of faces: A survey, Proc. IEEE, vol. 83, pp , [4] N. Ramanathan and R. Chellappa, Face Verification across Age Progression, IEEE Trans. on Image Processing, pp , [5] N. Ramanathan and R. Chellappa, Modeling Age Progression in Young Faces, CVPR, pp , [6] A. R.Martinez and R. Benavente, The AR face database, Computer Vision Center (CVC) Technical Report, Barcelona, Technical Report 24, [7] A. Georghiades, D. Kriegman, and P. Belhumeur, From few to many: generative models for recognition under variable pose and illumination, IEEE Transactions on Pattern Analysis and Machine Intelligence, pp , [8] M. Turk and A. Pentland, Face recognition using eigenfaces, In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, [9] M. Lyons, S. Akamatsu, M. Kamachi, and J. Gyoba, Coding facial expressions with Gabor wavelets, In 3rd International Conference on Automatic Face and Gesture Recognition, pp , [10] R. Gross, Face Databases, Handbook Of Face Recognition, Springer-Verlag, [11] P. J. Phillips, H. Moon, S. Rizvi, and P. J. Rauss, The FERET evaluation methodology for face-recognition algorithms, IEEE Transactions on Pattern Analysis and Machine Intelligence, pp , [12] W. Gao, B. Cao, S. Shan, D. Zhou, X. Zhang, and D. Zhao, CAS-PEAL large-scale Chinese face database and evaluation protocols, Joint Research & Development Laboratory Technical Report JDL-TR- 04-FR-001, [13] T. Sim, S. Baker, and M. Bsat, The CMU pose, illumination, and expression (PIE) database, In Proc. of the 5th IEEE International Conference on Automatic Face and Gesture Recognition, [14] T. Kanade, H. Saito, and S. Vedula, The 3D room: digitizing time-varying 3D events by synchronized multiple video streams, CMU Robotics Institute, Technical Report CMU-RI-TR-98-34, [15] Lanitis and Andreas, n.d., FG-NET Aging Database, Available: November, 17, [16] P. Ekman and W.V. Friesen, Facial Action Coding System, Consulting Psychologist Press, [17] L. G. Farkas, Anthropometry of the Head and Face, Raven Press, New York, [18] W. Horng, C. Lee and C. Chen, Classification of age groups based on facial features, Tamkang Journal of Science and Engineering, Vol. 4, No. 3, pp , [19] Geng X, Zhou ZH, Chen SF, Eye location based on hybrid projection function, Journal of Software, [20] I.T. Jolliffe, Principal Component Analysis, 2nd ed. Springer, [21] J. R. Beveridge, K. She, B. A. Draper, and G. H. Givens, A nonparametric statistical comparison of principal component and linear discriminant subspaces for face recognition, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp , December