Comparative Study of Lightweight CNN Architectures and FaceNet-based Transfer Learning for Face Recognition in Smart Attendance Systems

Abstract

Automated face recognition has emerged as a critical enabler of touchless attendance management, addressing growing demands for hygienic, contact-free identification in academic and professional environments. This paper presents a systematic comparative study of three custom-trained Convolutional Neural Network architectures — CNN-7, CNN-9, and CNN-11 — and a transfer-learning pipeline based on FaceNet (InceptionResnetV1 / VGGFace2) paired with Support Vector Machine (SVM), K-Nearest Neighbour (KNN), and Softmax classifiers. Experiments are conducted across three dataset configurations: (i) a controlled custom dataset of 1,890 face images from 30 subjects (15 male, 15 female), captured at nine illumination levels (−40% to +40% in 10% increments) and seven rotation angles (-90° to +90° in 30° increments); (ii) the publicly available PINS Face Recognition dataset comprising 50 celebrity subjects and 9,721 in-the-wild images; and (iii) a combined 80-class benchmark of 11,610 images formed by merging both sources. Among the custom CNN models, CNN-11 — comprising five convolutional blocks with Batch-Normalization and two Dropout regularisation layers — achieves the highest test accuracy of 100% on the custom dataset, followed by CNN-9 at 99.63% (F1 = 0.9963) and CNN-7 at 89.63%. The FaceNet + SVM pipeline attains perfect classification (100% accuracy, F1 = 1.0) on the isolated custom dataset and 99.66% accuracy (F1 = 0.9967) on the PINS dataset. On the combined 80-class benchmark, FaceNet + Softmax achieves 99.89% accuracy, with a five-fold cross-validation mean of 99.72% ± 0.04%, confirming robust generalisation across both controlled and unconstrained imaging conditions. All models are evaluated using accuracy, misclassification rate, sensitivity, specificity, precision, F1-score, and ROC-AUC. The system additionally incorporates a Duplicate Face Filter (DFF) module, an adaptive confidence-threshold real-time inference pipeline, and a desktop-based attendance monitoring application.