Human face localization and detection in highly occluded unconstrained environments

Abstract

Significant advancements have been achieved in the field of computer vision pertaining to the detection of human faces. This technological development holds great potential for a wide range of applications including but not limited to identification, surveillance and expression recognition. Unconstrained face identification has been significantly improved by the advancements in Deep Learning algorithms (DL). However, the presence of severe occlusion is an ongoing obstacle particularly when it obstructs a substantial section of the facial area, resulting in the absence of crucial facial characteristics. Furthermore, the limited availability of comprehensive datasets containing substantially obscured faces exacerbates the problem, impeding the efficacy of face detection programs. This study presents a new methodology, which incorporates an advanced occluded face detection (OFD) model, in order to enhance feature extraction and detection network. A dataset was developed specifically for training and testing the model. The new dataset includes faces with significant occlusion. The utilization of contextual-based annotation approaches improves the depiction of crucial facial characteristics. The OFD model exhibits exceptional performance and attaining a notable accuracy rate of 57.84%, a precision rate of 73.70% and a recall rate of 42.63%. These results surpass those achieved by alternative methods such as YOLO-v3 and Mobilenet-SSD. This study shows the capacity to make substantial progress in detecting occluded faces, hence offering the ability to make a positive influence on the domains of identification, surveillance and expression recognition.