An Attention-Enhanced Lightweight CNN Framework with MTCNN Detection and TripletEmbedding Recognition for Occlusion-Robust Automated Attendance from Surveillance Video

Abstract

Manual and card/barcode-based attendance recording remains slow, error-prone, and vulnerable to proxy marking, motivating fully automated, camera-based alternatives for schools and organizations. This paper proposes an Attention-Enhanced Lightweight CNN framework that couples MTCNN multi-scale face detection with a CBAM (Convolutional Block Attention Module) augmented MobileFace-style backbone trained under triplet loss to produce compact, discriminative 128-dimensional face embeddings from surveillance video. Enrolled identities are matched via cosine similarity against a reference embedding gallery, and a temporal multi-frame voting stage consolidates predictions across consecutive frames to suppress transient misdetections caused by pose change, partial occlusion, or motion blur.