|
Age and Gender Prediction using Deep cnns and Transfer Learning3
Methodology
3.1
Dataset
In this paper, we use the UTKFace dataset [2] (aligned and cropped) consists of over
20,000 face images with annotations of age, gender, and ethnicity. It has a total of 23708
images of which 6 were missing age labels. The images cover large variations in facial
expression, illumination, pose, resolution and occlusion. We chose this dataset because
of its relatively more uniform distributions, the diversity it has in image characteristics
such as brightness, occlusion and position and also because it involves images of the
general public.
Some sample images from the UTKFace dataset can be seen in Fig. 1. Each image
is labeled with a 3-element tuple, with age (in years), gender (Male-0, Female-1) and
races (White-0, Black-1, Asian-2, Indian-3 and Others-4) respectively.
Fig. 1. Sample images from the UTKFace dataset.
For both our approaches (custom CNN and transfer learning based models), we used
the same set of images for training, testing and validation, to have standardized results.
This was done by dividing the data sets into train, test and validation in 80 : 10 : 10
ratios. This division was done while ensuring that the data distribution in each division
remains roughly the same, so that there is no distribution mismatch while training and
testing the models. The Table 1 and Table 2 show the composition of training, valida-
tion and test data with respect to gender and age respectively.
4
Table 1. Composition of sets by gender
Gender
Training
Validation
Test
Total
Male
9900
1255
1234
12389
Female
9061
1115
1137
11313
Total
18961
2370
2371
23702
Table 2. Composition of sets by age
Age Group
Training
Validation
Test
Total
0-10
2481
303
278
3062
11-20
1222
150
158
1530
21-30
5826
765
753
7344
31-40
3618
462
456
4536
41-50
1767
223
254
2244
51-60
1858
214
226
2298
61-70
1057
137
122
1316
71-80
577
57
65
699
81-90
413
45
46
504
91-100
114
11
12
137
101-116
28
3
1
32
Total
18961
2370
2371
23702
|
| |