Optimizing occlusion detection in facial images using a combined autoencoder and SVM framework

Abstract

Over time, face recognition systems have become popular and are widely used in varied applications today. However, these systems have significant challenges especially when faced with occlusions which can greatly reduce the identification and verification accuracy of such systems. This study proposes a hybrid model that integrates Autoencoders and Support Vector Machines (SVM) for detecting occlusions in face recognition. In this model two stages are trained and evaluated using AR face database. The first stage comprises of an autoencoder trained to re construct facial images under occlusion with high reconstruction errors suggesting possible anomalies. On the other hand, the second stage uses SVM to distinguish between occluded and non-occluded on these anomalies. The purpose of this method is to guarantee an efficient recognition rate with good performance even under conditions when there is some kind of obstruction over the subject’s face. Experimental results showed that the proposed approach was useful, having obtained a 94.75% success rate, optimized through parameter tuning. Therefore, this work suggests that integration of autoencoders along with SVMs can improve occlusion detection and hence, face recognition systems’ performance while also overcoming its limitations like class imbalance.