Age and Gender Prediction using Deep CNNs and Transfer Learning

TL;DR

This paper investigates automatic age and gender prediction from facial images using deep CNNs and transfer learning, comparing custom and pre-trained models, and analyzing their effectiveness for different classification tasks.

Contribution

It introduces a comprehensive comparison of custom CNNs versus transfer learning with pre-trained CNN architectures for age and gender prediction from facial images.

Findings

Pre-trained CNNs outperform custom CNNs in most cases.

Linear regression on extracted features can outperform end-to-end CNN training.

Transfer learning with VGGFace and VGGFace2 improves prediction accuracy.

Abstract

The last decade or two has witnessed a boom of images. With the increasing ubiquity of cameras and with the advent of selfies, the number of facial images available in the world has skyrocketed. Consequently, there has been a growing interest in automatic age and gender prediction of a person using facial images. We in this paper focus on this challenging problem. Specifically, this paper focuses on age estimation, age classification and gender classification from still facial images of an individual. We train different models for each problem and we also draw comparisons between building a custom CNN (Convolutional Neural Network) architecture and using various CNN architectures as feature extractors, namely VGG16 pre-trained on VGGFace, Res-Net50 and SE-ResNet50 pre-trained on VGGFace2 dataset and training over those extracted features. We also provide baseline performance of various machine learning algorithms on the feature extraction which gave us the best results. It was observed that even simple linear regression trained on such extracted features outperformed training CNN, ResNet50 and ResNeXt50 from scratch for age estimation.