Face Recognition Using $Sf_{3}CNN$ With Higher Feature Discrimination

TL;DR

This paper introduces Sf3CNN, a framework combining 3D ResNet and A-Softmax loss to improve face recognition accuracy in videos, achieving over 99% accuracy on a challenging database.

Contribution

The paper proposes Sf3CNN, a novel framework that leverages 3D ResNet and A-Softmax loss for enhanced video-based face recognition.

Findings

Achieved 99.10% accuracy on CVBL video database.

Outperformed previous methods with 97% accuracy using 3D ResNet.

Effectively captures spatial and temporal features for face recognition.

Abstract

With the advent of 2-dimensional Convolution Neural Networks (2D CNNs), the face recognition accuracy has reached above 99%. However, face recognition is still a challenge in real world conditions. A video, instead of an image, as an input can be more useful to solve the challenges of face recognition in real world conditions. This is because a video provides more features than an image. However, 2D CNNs cannot take advantage of the temporal features present in the video. We therefore, propose a framework called $Sf_{3}CNN$ for face recognition in videos. The $Sf_{3}CNN$ framework uses 3-dimensional Residual Network (3D Resnet) and A-Softmax loss for face recognition in videos. The use of 3D ResNet helps to capture both spatial and temporal features into one compact feature map. However, the 3D CNN features must be highly discriminative for efficient face recognition. The use of A-Softmax loss helps to extract highly discriminative features from the video for face recognition. $Sf_{3}CNN$ framework gives an increased accuracy of 99.10% on CVBL video database in comparison to the previous 97% on the same database using 3D ResNets.