KInsight: a Robust Framework for Masked Face Recognition

Abstract

The pandemic and other environmental conditions have increased the use of masks as a precautionary measure. It is an effective way to protect ourselves from viruses/pollution/ and other health-affecting environmental factors. However, for smart devices such as smartphones with face locks, attendance systems, and smart surveillance cameras with enabled face recognition, these masks raised another challenge of masked face recognition. Masked face recognition is an increased subset challenge of the standard face recognition problem as they lack facial features. The masked images are occluded, making the structure and the facial features of the non-occluded region of importance. This paper presents a novel two-fold approach KInsight (K Nearest Neighbor-based Insight Face algorithm) for masked face detection using antelopev2, which uses a RetinaFace detection algorithm and a ResNet100 Convolutional Neural Network for face detection and embedding generation. Further, we propose to use a KNN classifier for masked face recognition. Several experiments have compared the proposed scheme’s performance with important research contributions. Experimental results show that the scheme significantly outperforms several benchmark approaches with an accuracy of around 98.5%.

Keywords:
Face Recognition; Deep Neural Network; Face detection

HIGHLIGHTS

An integrated scheme to improvise masked face detection & recognition is proposed

The scheme Integrate Dlib & Retina face to detect oversized & low-resolution images

Results shows an accuracy of 98.5 % over benchmark dataset.

GRAPHICAL ABSTRACT