Age-gender estimation

This paper is a product of KAIST IVY Lab. and Genesis Lab through the IITP 2018 R&D Voucher Project.

Original Paper

Adversarial Spatial Frequency Domain Critic Learning for Age and Gender Classification

Proposed Method

The main idea in this paper is to combine age and gender information, which predominantly appears in the spatial frequency domain, into the generated image. Another technique attempted is training by calculating the loss alternately between age and gender during training. Details are as follows.

Fig. 1. Age and Gender Classification Network Using Adversarial Spatial Frequency Domain Critic

1. Encoder-Generator

The encoder-generator is similar to DCGAN. The encoder is a CNN network that extracts features from the input real image, and the generator creates a fake image using the input value. The difference from DCGAN here is that the output value of the encoder is combined with the age and gender labels and used as the input value for the generator. The generator, receiving the age and gender information, tries to use it to generate a face image with the added age and gender attributes.

2. Adversarial Spatial Frequency Domain Critic

The adversarial spatial frequency domain critic plays a role in reducing the appearance-wise discriminability and noise of the generated image while preserving the characteristics of age and gender.

As shown in Fig. 2. (a) and (b), when the public image dataset was classified by age and gender classes and the average values were calculated, and the slopes of the CNN output layer activations were investigated for these, different regions were activated for each. As shown in Fig. 2 (c) and (d), in the activation classified by age, textures such as wrinkles stood out prominently, and in the activation classified by gender, facial landmarks such as eyes, nose, and mouth stood out prominently.

Fig. 2. Average images and activations by class

The characteristics of age and gender appeared predominantly in different spatial frequency domains, as shown in the images in Fig. 3 (a) and (b), which are the result of multiplying the activation of each class by the Fourier-transformed image.

Fig. 3. (a) is the result of multiplying the spatial frequency of Fig. 2 (a) by Fig. 2. (c), (b) is the result of multiplying the spatial frequency of Fig. 2. (b) by Fig. 2. (d)

Fourier transform decomposes a function of time or space into time or spatial frequency components, and is also used for the purpose of filtering specific characteristics by selecting only the desired frequencies.

Using these characteristics of spatial frequency, a mask is generated to preserve the characteristics of age and gender and reduce other characteristics. The mask is created differently for age and gender respectively, and it plays a role in reducing characteristics by multiplying by 1 in the spatial frequency domain where age and gender characteristics are dominant, while multiplying by a constant between 1 and 0 in other spatial frequency domains.

Equation 1. Critic mask equation

Equation 2. Critic loss function equation

3. Multi-Task Discriminator

The discriminator proposed in this paper plays two main roles. Similar to GAN, it determines the authenticity of the classified image and classifies age and gender. The loss function is also calculated for each role, and age and gender are calculated separately. Age is trained to be classified into 8 classes using the cross entropy loss function.

4. Alternating Learning

Although training to classify age and gender is conducted in the same network, the strategy is to proceed with training alternately. As in Algorithm 1, when the encoder-generator training to reduce the encoder-generator loss and the critic loss for gender, and the training to reduce the discriminator's loss for gender classification are finished, the encoder-generator training to reduce the encoder-generator loss and the critic loss for age, and the training to reduce the discriminator's loss for age are conducted. Training alternates like this in every epoch.

Algorithm 1.

Experimental Results

Experiments were conducted using the Adience benchmark dataset and the LFW dataset. As a result of comparing each data targeting hand-crafted methods and CNN-based methods, the method using the mask introduced in this paper showed higher accuracy than other methods. Even the accuracy in classifying age showed superior performance compared to other methods, even without using the mask introduced in this paper.

Conclusion

It was confirmed that the spatial frequency domain critic network and the alternating training strategy for age and gender proposed in this paper achieve superior performance in age and gender classification tasks than other compared methods. By filtering specific regions of the spatial frequency domain, age and gender information were further preserved in the generated images, and the ability to classify age and gender could be further enhanced. with the alternating training strategy.

References

[5] Eidinger et al., “Age and Gender Estimation of Unfiltered Faces,” IEEE TIFS 2014.

[11] Levi et al., “Age and Gender Classification Using Convolutional Neural Networks,” CVPRW 2015.

[12] Hsieh et al., “Multi-Task Learning for Face Identification and Attribute Estimation,” ICASSP 2017.

[20] Hassner et al., “Effective Face Frontalization in Unconstrained Images,” CVPR 2015.