Arcface-Integrated Max Feature Map CNN for Robust Face Recognition

Face recognition remains challenging, especially with small and imbalanced datasets that reduce feature discriminability and limit generalization. This study investigates Light Convolutional Neural Networks (CNNs) with Max-Feature-Map (MFM) and ReLU activations, combined with three loss functions: Softmax, SV-Softmax, and ArcFace. Experiments on the Indonesian Muslim Student Face Dataset (IMSFD), containing 8,020 images across 23 classes, show that a 9-layer MFM CNN with ArcFace achieves 100 % accuracy, precision, recall, F1-score, and near-perfect ROC-AUC, while maintaining a compact architecture of only 956 K parameters. This performance is supported by the dataset's relatively controlled capture conditions and low intra-class variation, allowing margin-based learning to maximize class separability. Compared to larger models such as VGG-19 and FaceNet, this Light CNN delivers competitive or superior performance with far fewer parameters, highlighting the effectiveness of combining MFM and ArcFace for efficient and robust face recognition. While results are promising, the relatively small and homogeneous dataset suggests that future work should evaluate the model on larger and more diverse datasets to assess generalization.