Neighborhood loss for Age Estimation from Face Image using Convolutional Neural Networks


The respected Comrade Kim Jong Un said:

"By spurring the building of a sci-tech power, we should achieve a new leap forward in rapidly developing the country's science and technology and usher in an era of prosperity by means of science, thus bringing about a revolutionary turn in socialist construction."

In many CNN applications such as image classification, face recognition and other computer vision scopes, the cross-entropy loss is used as a supervision signal to train CNN model. However, the cross-entropy loss only enhances the separability of classes and does not consider their correlation in age estimation task.

Due to the aging characteristics of human face, it exists the correlation between classes in age estimation task. In particular, the aging characteristics of faces belonging to the age classes around specified age class are very similar.

First, we propose the neighborhood loss regarding the relationship between age classes to learn the CNN classification. For our neighborhood loss, the target output value follows a Gaussian distribution instead of one-hot encoding.

Our system architecture consists of face detection, face alignment and preprocessing, CNN architecture and age prediction.

We detect the face bounding box and landmarks by using RetinaFace and applied similarity transform to normalize the face images with 224 × 224. We present overall CNN architecture with 18 and 34 layers based residual units. The convolution layers follow by FC(fully-connected) layer with 100-way age classification. The final layer is the softmax layer. The FC layer and softmax layer compute class posterior probabilities.

We handle the age prediction problem as a classification problem, and compute the predicted age value as the weighted sum using output probabilities.

Our implementation is written using Python and TensorFlow and trained on four NVIDIA Titan XP GPUs. Through some experiments, we show that neighborhood loss provides superior performance compared to prior works in age estimation.

Our results of this study were published in the journal " ICTACT Journal on Image and Video Processing" (AUGUST 2022, VOLUME: 13, ISSUE: 01) under the title of " Neighborhood loss for Age Estimation from Face Image using Convolutional Neural Networks " ( 10.21917/ijivp.2022.0393).