HISANet: A Channel Attention-Based Siamese Network for Robust Face Verification

Abstract

Face verification in unconstrained environments remains a challenging task due to significant variations in illumination, pose, expression, and occlusion, which can degrade the stability of learned representations.Although deep metric learning approaches have demonstrated strong performance on large-scale datasets, their effectiveness often diminishes in data-constrained or heterogeneous scenarios, where global feature embeddings may capture non-discriminative or noisesensitive patterns.To address this limitation, this paper introduces the Hybrid Invariant-Siamese Attention Network (HISANet), a lightweight Siamese architecture augmented with a channel attention mechanism based on the Squeeze-and-Excitation principle.The proposed model adaptively recalibrates channel-wise feature responses, enabling the network to emphasize identity-relevant information while suppressing less informative variations.A shared convolutional backbone is used to generate compact embeddings, and similarity between image pairs is computed using scaled cosine similarity within a binary classification framework.The network is trained end-to-end using a binary cross-entropy objective and evaluated on the Labeled Faces in the Wild (LFW) dataset following the standard verification protocol.A comprehensive evaluation, including accuracy, area under the ROC curve, equal error rate, and qualitative analysis, demonstrates that the integration of channel attention improves feature selectivity and reduces overfitting.HISANet achieves a validation accuracy of 75.83% and an equal error rate of 20.13%, indicating that attention-based feature recalibration can enhance robustness while maintaining computational efficiency.