FACIAL RECOGNITION METHOD UNDER ARBITRARY VIEW

Abstract

The article addresses the challenge of utilizing facial recognition technology in industrial applications. Despite the integration of this technology, there are open issues such as verification and identification of individuals from different poses. The absence of proper research in facial recognition in videos, especially in surveillance systems using snapshots from various Points of View (POV), is a particular concern. These challenges are emphasized in the context of using photographs taken frontally and from the right profile, traditionally collected by the police.

To address these issues and fill the research gap, a new approach is proposed in the form of the Face Recognition from Mugshots Database (FRMDB). This database comprises 28 snapshots and 5 videos taken from different angles for 39 unique subjects. The main objective of FRMDB is to analyze the impact of using snapshots from various perspectives on the accuracy of facial recognition in surveillance video frames.

To evaluate the effectiveness of FRMDB and make comparisons with existing data, accuracy tests were conducted using two deep neural networks (CNNs), namely VGG16 and ResNet50. They were pre-trained on the VGGFace and VGGFace2 datasets for extracting facial features. A comparative analysis of the results was conducted using data from the existing research, specifically the Surveillance Cameras Face Database (SCFace).

The results underscore that the subset of snapshots, including frontal and right profile images, exhibits the lowest accuracy among the tested variations. This indicates the need for additional research to determine the optimal number of snapshots for effective facial recognition in surveillance video frames.