Enhanced contextual based deep learning model for niqab face detection

Abstract

Human face detection is one of the most investigated areas in computer vision which plays a fundamental role as the first step for all face processing and facial analysis systems, such as face recognition, security monitoring, and facial emotion recognition. Despite the great impact of Deep Learning Convolutional neural network (DL-CNN) approaches on solving many unconstrained face detection problems in recent years, the low performance of current face detection models when detecting highly occluded faces remains a challenging problem and worth of investigation. This challenge tends to be higher when the occlusion covers most of the face which dramatically reduce the number of learned representative features that are used by Feature Extraction Network (FEN) to discriminate face parts from the background. The lack of occluded face dataset with sufficient images for heavily occluded faces is another challenge that degrades the performance. Therefore, this research addressed the issue of low performance and developed an enhanced occluded face detection model for detecting and localizing heavily occluded faces. First, a highly occluded faces dataset was developed to provide sufficient training examples incorporated with contextual-based annotation technique, to maximize the amount of facial salient features. Second, using the training half of the dataset, a deep learning-CNN Occluded Face Detection model (OFD) with an enhanced feature extraction and detection network was proposed and trained. Common deep learning techniques, namely transfer learning and data augmentation techniques were used to speed up the training process. The false-positive reduction based on max-in-out strategy was adopted to reduce the high false-positive rate. The proposed model was evaluated and benchmarked with five current face detection models on the dataset. The obtained results show that OFD achieved improved performance in terms of accuracy (average 37%), and average precision (16.6%) compared to current face detection models. The findings revealed that the proposed model outperformed current face detection models in improving the detection of highly occluded faces. Based on the findings, an improved contextual based labeling technique has been successfully developed to address the insufficient functionalities of current labeling technique. Faculty of Engineering - School of Computing183http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:150777 Deep Learning Convolutional neural network (DL-CNN), Feature Extraction Network (FEN), Occluded Face Detection model (OFD)