Deep Residual Learning for Human Identification Based on Facial Landmarks

Abstract

The face detection and recognition are still challenging research areas, since up to date there is no accurate integral model that works in every situation. As a result, the focus has been shifted to Convolutional Neural Networks (CNNs) and fusion techniques with the hope of better solution. The CNNs have enhanced the state-of-the-art of the human facial identification. However, the CNNs are not easy to train due to degradation problem called (gradient vanishing) when the network depth increased, so there is a need of residual network to solve this problem by going deeper without losing the gradient. In this paper, a pre-trained deep residual network for features extraction and ensembles of classifiers are implemented and facial landmarks are extracted and passed to the pre-trained model. Support Vector Machine (SVM) and random forest classifiers are fused at decision level using weighted and majority voting techniques. The experimental results conducted on ORL database show an excellent mean accuracy rate of about 100%. The accuracy rate of about 100% was achieved on LFW dataset with a minimum 70 facial images per person, and 99% with a minimum 10 facial images per person.