FIAC-Net: Facial Image Attractiveness Classification Based on Light Deep Convolutional Neural Network

Automatic facial beauty assessment is an emerging research area with various applications. Modeling beauty objectively is an extremely challenging task, mainly due to the lack of a globally approved face beauty assessment procedure. Secondly, the scarcity of an adequate facial image dataset dedicated to facial beauty and attractiveness prediction and the way of representing beauty-related features. Therefore, facial beauty prediction is appropriate for data-driven and learned features techniques such as deep learning rather than hand-crafted features and machine learning approaches. Recently, researchers have employed deep convolutional neural networks (DCNN) aspects and transfer learning to overcome the size of the small dataset and the overfitting problem. However, large CNN designs are still challenging to implement on small devices with limited hardware resources since they contain millions of parameters, making them computationally demanding. This paper provides a light deep convolutional neural network (LDCNN)for facial images attractiveness assessment of both genders from scratch called facial image attractiveness classification network (FIAC-Net), with a reduced number of parameters compared to the pre-trained networks to make it possible to deploy deep neural networks on compact hardware. Soft discrete labels are created for the continuous beauty scores of both SCUT-FBP and SCUTFBP5500 datasets with augmented data to enhance the learning ability to be adapted to the attractiveness classification of the CelebA dataset. The results reveal that the proposed network architecture achieved good results.