Potential of Robust Face Recognition from Real-Time CCTV Video Stream for Biometric Attendance Using Convolutional Neural Network

摘要

Face recognition is one of the most bothersome research issues in security systems due to various challenges like constantly changing poses, facial expressions, lighting conditions, and resolution of the image. The wellness of the recognition technique firmly depends on the accuracy of extracted features and also on the ability to deal with the low-resolution face images. The mastery to learn accurate features from raw face images makes deep convolutional neural networks (DCNNs) a suitable option for facial recognition. The DCNNs utilizes Softmax for evaluating model accuracy of a category for associate degree input image to create a forecast. However, the Softmax probabilities do not depict the real representation of model accuracy. The main aim of this paper is to maximize the accuracy of face recognition systems by minimizing false positives. The complete procedure of building a face recognition prototype is defined very well. This prototype consists of many vital steps built using most advanced methods: CNN cascade for detection of face and HOG for generating face embeddings. The primary aim of this analysis was the sensible use of those developing deep learning techniques for face recognition work, because of the reason that CNNs give almost accurate results for huge datasets. The proposed face recognition prototype can be used together with another system by making some minor changes or without making any changes as an assisting or a primary element for surveillance functions.