Triple-Stream Attention Network (TSAN): A Multi- Phase Deep Learning Framework for Robust Facial Recognition

Abstract: Facial recognition has already formed an essential
component of present-day security and surveillance systems, as
well as human-computer interaction, but robust recognition in
the uncontrolled environment is a major issue owing to changes
in pose, illumination, occlusions, and facial expressions. To solve
these issues in this study, a multi-phase deep learning model
Triple-Stream Attention Network (TSAN) is proposed to combine
local, global, and structural facial contexts to enhance recognition
accuracy and resilience. TSAN uses three stream architecture
that includes a convolutional neural network(CNN) uses local
texture features, a Vision Transformer uses global context, and a
Graph Convolutional Network(GCN) uses geometric structure
based on facial landmarks. To optimize the features
representation and classification, hierarchical attention fusion
and hybrid loss function with Cross-Entropy and Center Loss are
used. This model was trained and tested on the Labeled Faces in
the Wild (LFW) dataset, and had a recognition accuracy of
99.2% which surpasses the state-of-the-art accuracy of FaceNet,
ArcFace, SphereFace and CosFace. The findings prove that
TSAN is effective in capturing complementary information and
achieves a strong performance across different circumstances
thus can be used in a real world context in facial recognition
devices.
Keywords: Facial Recognition, Deep Learning, Triple-Stream
Attention Network, Vision Transformer, Graph Convolutional
Network, Hierarchical Attention, LFW Dataset