An Approach of Detecting the Age of a Human by Extracting the Face Parts and Applying the Hierarchical Methods

Age of a person can be estimated in several ways by the usage the several advancements in different fields. Popularly used ways are categorised as physical activity, medical records and tests, and by processing the images.

The Walking style, handwriting are examples of physical activity, With the change in age, the pattern of walking changes. Hence walking factor can be used for age prediction [1]. Handwriting also aids in detecting the age, gender and nationality [2].

DNA methylation tests, MRI, and EMR, are sources of tests used for age prediction. It is difficult to predict the age because most of the people are not seems as their age. Chronological age is the age that a person alive while the biological age is the person seems. Biological age can be predicted with blood biomarkers [3]. Overall electronic medical records are sources of biological age. It is possible to estimate the chronological age from EMRs using deep neural networks [4]. By proper feature selection, matching younger age faces with the older one is possible [5].

With the increase in age, few features of face changes. By extracting those, one can estimate the age [6]. Ageing affects many tissues. Hence it can be estimated by considering the different tissues and DNA methylation [7]. The epigenetic clock is the test based on DNA methylation levels which are extensively used in forensic age prediction. MRI tests can aid in brain age prediction [8, 9]. MRI data of hand bones, clavicle bones, and wisdom teeth can help assist the age [10].

The processing of images doesn’t require any physical tests. it is possible to design the age estimator by training the human faces with age difference information [11]. There were different age databases [12] available for research in age. In this paper, all the age images are collected from Kaggle.

The age can be estimated by sorting the facial images as per age in time plane and extracting the features [13]. It is difficult to predict the exact age because ageing is a slow process, and there is not much feature difference between the adjacent ages. Hence while learning from a set of images related to age. It should not only learn that age but also contributes to learning the adjacent ages [14]. Extraction of attributes of the face with age progression helps in reliable age estimation [15]. Estimating with face attributes in camera images is a challenging task, the dynamic features extraction can make this task easier [16]. Instead of using holistic features, discriminative local features can aid in better age estimation [17].

Facial landmarks assist in improving the efficiency of age prediction [18]. Some of the effects like Lightning, pose etc., may reduce the chances of extracting the features. 2D to 3D modelling reduces the effect of Lightning and pose variations [19]. Computational intelligence attained an excellent accuracy of age estimation in recent years [20].

SVM is used extensively for age regression [21]. Instead of using SVM, Gaussian process can also be used for the age estimation [22]. Some methods like the Manifold learning method can also be used for age estimation [23].

A common neural network is used for all the classifications. The neural network is selected by performing the operations on two categories of images- below 40 and above 40 age. Because of craniofacial growth, there would be considerable changes in the faces of kids and adults. Kids skull is small, and hence their face appeared to be small. The newborn baby cannot up his head with his neck. The tiny face of kids is just one-third of their skull, but adult’s face is half of their skull. The chins are smaller for kids. Hence their faces are visible to be round. During the child phase, the face changes are more with age, and it continues in decreasing order with age. When transformed into adults, moustache and beard will develop for males. While changing from adult age to older age, the skin becomes loose, and wrinkles in the face will be formed. The whites of eyes of older people may change to yellow due to the exposure of dust, wind and UV light for many years. By doing face image processing well, the machine can able to predict the age more accurately. The main aim of the Block1 is to extract the features. The minimum L value considered is 32, which means at least 32 filters are used for extracting the features in the face. The images are converted from BGR to greyscale before training. All the faces are different shaped photos some are landscape, portrait. All are converted to portrait and normalized to a common size. As the size increases, the clarity would be more but need to select the same size for all. It is found that 100x100 is enough to extract the features. Figure 1 presents an illustration of loss in clarity of images with size.

Hence all the images used for training are normalized to 100x100 pixels. The training data must be properly balanced. For each age up to 60, the training images used is 50 for training and above 60 age images are 1030. So for below 40 category, the total images are 40x50=2000. And for above 40, the total images are 20x50+1030=2030. 70% of images are used for training and 30% of images are used for validation. If one category of data trained first and another category next which makes not proper training. Hence the images were shuffled before training. It is difficult to work with images with the change in neural network Hence the training images are converted to pickles for easy access.

Figure 4 illustrates various neural networks that are obtained with different combinations of L, C and D values. Where L is used for convolution layer size, C for Number of convolution layers and D for Number of Dense layers. Block 1 and Block 2 of Figure 4 are illustrated in Figure 2 and Figure 3. The accuracies of various combinations of these parameters along with epochs are illustrated in Table 1. The training accuracy < 100% at the epochs 10. The possible reason would be underfitting or improper training. Since training accuracy>validation accuracy, it is not due to underfitting. The training can be improved with an increase in epochs, and it is found that by improving epochs to 100, the training accuracy reaches 100%.

For 10 Epochs of training, keeping L and C constant, the training accuracy increases with increase in D in most of the cases. In the case of L=64 and C=1, the training accuracy is reached up to 97%. But the improvement in training accuracy is achieved well by increasing epochs. The training accuracy reaches its maximum for the epochs from 50 to 100. But there is no much effect on validation accuracy with the increase in epochs.

Considering all of these, L=32, C=3 and D=3 is selected as a neural network for all the operations. Let the neural network be N. Epochs is increased to 200 for N, the validation accuracy achieved is 84.4 which is not a considerable improvement. Since validation accuracy < training accuracy, there is a chance of overfitting. Batch normalization can regularize this. Besides, it can increase the speed of training and can decrease the importance of initial weights. Dropout also can reduce overfitting. Hence Batch normalization and a dropout of 0.25 is added before the sigmoid activation layer of Figure 4. The accuracy now achieved is 80.5 and the same layers are added after each of Block 2, the accuracy obtained is 80.3. Addition of these layers for 200 epochs doesn’t affect much. Same is trained with 100 epoch which results in 82.2 for layers addition for each of Block 2. 81% of accuracy is achieved for the addition before sigmoid activation function. From the above results, the neural network N with 100 epochs is selected for all the pieces of training.

1.png

200x200                    100x100                     50x50                         30x30

Figure 1. Effect of image quality at different sizes

2.png

Figure 2. Block 1

3.png

Figure 3. Block 2

4.png

 Figure 4. Neural network

Table 1. Selection of neural network

To improve accuracy, a hierarchical method is designed. The hierarchy method consists of various trained models which is illustrated in Figure 5. First, the hierarchy method predicts whether the age is below 40 or above 40 with the aid of 40 na 40 trained model. If below 40 then it is checked with the models trained with 20 age gap. In this case 20n40 followed by 10 and at the end 5 age gap. Hence the highest priority is 40 followed by 20,10, and 5 age gaps. The accuracies achieved for the various age gaps are as follows-

40: 40na40 – 81%

20: 20n40 – 82.6% and (40-60)an(A60) – 75.5%

10: 10n20 - 81.37%, 30n40 – 68.2% and 50n60 – 57.19%

05: 5n10 – 73.2%, 15n20 – 71.24%, 25n30 – 60.13%, 35n40 – 50.66%, 45n50 – 62.89%, 55n60 – 61.63%.

5.png

Figure 5. Block diagram of the Hierarchical method

Table 2. Hierarchical method

The notation xnax represents the trained model having two categories – belowx, above x. xny represents the trained model having two categories – (above [2x-y] and below x), (above x and below y). The notation (x-y)an(Ay) represents the trained model having two categories – abovex and belowy, above y.

Table 2 illustrates the accuracies achieved at different age ranges. It is not easy to find the age within the age-gap of 5. Hence the accuracy of 5 age gap is less compared to 20 age gap. Nevertheless, for the above 60age category, the ages are from 60 to 100. Hence it becomes easier for the model to find the person having the age above 60 and so the accuracy for above 60 is vast compared to 5 age gap.

Table 3. Accuracies of normal training

The extracted face parts are trained with the Neural network N. The face part wise accuracy for the age gaps 20,10 and 5 considering all the ages and along with the combination of all face parts are presented in Table 3. The final accuracy FA is better than the normal accuracy achieved in section III but less than the accuracy achieved in section IV.

FA represents combination of all face part models.