ROBUST FACIAL RECOGNITION UNDER OCCLUSION

Abstract

The face detecting methodologies and frameworks are getting degraded due to outbreak of novel coronavirus, which resulted in the rise of face masks on all human. The presence of face masks on facial landmarks makes them more occluded and it gets difficult to detect the face properly as it is covering most of the important facial features such as nose and lips. Occlusion in face detection is defined by the angle of the face, image illumination, shadows, etc., but now face mask is a very crucial element to cause high level of occlusion. Frameworks sensitive to this occlusion problem fails to perform efficiently and produces inaccurate experimental results. To overcome this problem an improved and efficient CNN framework is introduced which performs notably better under occluded conditions. The introduced framework processes by using three separate layers of CNN models such as P-net, R-net and O-net for face detection with low computational time and complexity. MTCNN processes by resizing the images to detect different sizes of faces and uses and feeds the scanned images between all the neural networks by performing non-maximum suppression to eliminate the false-positive cases at every step to increase the accuracy and speed. The testing and generation of results are produced using MaskedFace-Net [22] and FFHQ [20] datasets.