An Evaluation of the Performance of Convolution Neural Network and Transfer Learning on Face Gender Recognition

Abstract

Gender detection has become an important component of the biometric system and protection. It is used to recommend age-appropriate content and to broaden the reach of multilayer advertising and promotions, among others. Many challenges emanate from using contemporary machine learning algorithms for face gender recognition, ranging from low detection accuracy and overfitting. Most recognition tasks, including object identification, segmentation, and classification, are based on convolutional neural networks. The human Face is a nonlinear manifold in which slight input changes can result in the invariance of the output. This prompts the need for continuous research to achieve more accurate recognition, especially with experimenting on ground truth data. This research compares the performance of CNN and TL on faces acquired in real-time scenarios to detect gender. The empirical result based on the study demonstrates that TL outperforms CNN in terms of prediction accuracy, with TL having 95.83% and CNN having 83.11%, and also in terms of recognition time, with TL having 0.23secs and CNN having 0.98 secs, amongst other evaluation metrics. The novelty of the research is in the recognition time, which is not a metric commonly reported in the literature. The work demonstrates the effectiveness of using pre-trained models based on the dataset used and evaluation metrics for the experiment. Keywords: Gender detection· CNN· Transfer learning· Face gender recognition