Deep Learning Feature Extraction Architectures for Real-Time Face Detection

• Published in: SN computer science Vol. 4; no. 5; p. 645
• Main Authors: B, Ravi Teja, D, Mythili, Duvva, Laxmiprasanna, Bethu, Srikanth, Garapati, Yugandhar
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 01.09.2023; Springer Nature B.V
• Subjects: Accuracy; Activity recognition; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Datasets; Deep learning; Enabling Innovative Computational Intelligence Technologies for IOT; Face recognition; Feature extraction; Image enhancement; Information Systems and Communication Service; Likelihood ratio; Machine learning; Neural networks; Original Research; Pattern Recognition and Graphics; Real time; Regression models; Software Engineering/Programming and Operating Systems; Surveillance systems; Vision; Feature extraction; CNN; Deep Learning
• ISSN: 2661-8907; 2662-995X; 2661-8907
• DOI: 10.1007/s42979-023-02023-5

A Video Surveillance system can be used for a variety of purposes, including protection, secure data, crowd flux analytics and congestion analysis, individual recognition, anomalous activity detection, and so on. Video Surveillance systems play the key role in the human detection using the face features extraction. It helps in many applications like terrorists attack, thief identifying by detecting the face of the person but mostly failed in real-time aspect. In this context, we propose a method that significantly aids in the extraction and learning of features. To reduce the face recognition error, we use a bounding box regression model. To train the features, we utilized a CNN-based feature learning model with log-likelihood ratio calculations between inter- and intra-features. To increase the quality of video frames, we used a histogram redistribution image enhancement technique. Finally, a Background Subtracted Faster RCNN for video-based face recognition (BSF-RCNN-VFR) is used to discriminate the groups of detected faces. A comprehensive experiment is carried out on their datasets to demonstrate that the proposed solution performs better, and we compared the existing models with proposed models. We achieved 94.2 accuracy percentage. In this paper, the CNN models like AlexNet, ResNet and datasets like UADFV, Celeb-DF, FF++, DFDC, etc., accuracies are compared.