Real-Time Gender Detection Using Convolutional Neural Networks with Viola–Jones Face Detection
TL;DR
A framework for automatic gender classification based on facial features in real time is presented, combining the Viola-Jones object detector with a lightweight Convolutional Neural Network (CNN), employs cross correlation for feature extraction and consists of convolution and subsampling layers.
Abstract
Face Gender Recognition can be considered a branch of computer vision that is used in various domains like biometrics, surveillance, interaction between humans and computers, and intelligent security systems. In this work, a framework for automatic gender classification based on facial features in real time is presented, combining the Viola-Jones object detector with a lightweight Convolutional Neural Network (CNN). The four-layer CNN employs cross correlation for feature extraction and consists of convolution and subsampling layers. The methodology followed includes preprocessing, data augmentation, normalization, feature extraction, training and testing the framework using the Kaggle Gender Classification Dataset and 80:20 train-test split. Results achieved include 98.75% and 99.38% accuracy on SUMS and AT&T databases, respectively. Processing rate of 32 x 32 facial images is less than 0.27 ms with more than 3,700 images/s processing and convergence in less than 20 epochs. Using OpenCV library, the framework also supports webcam-based prediction in real-time with more than 99.99% confidence levels.