Attention-Guided EfficientNet for Precise Identification in Surveillance Images

This paper proposes AG-EfficientNet, combining EfficientNet-B0 with Convolutional Block Attention Modules (CBAM), multi-scale feature fusion, and hybrid Softmax-Triplet optimization for face identification in degraded surveillance imagery. The authors report evaluation on the LFW and SCFace datasets, with 98.2% identification accuracy, 97.9% precision, 97.6% recall, 97.7% F1-score, and 0.99 ROC-AUC. The abstract says the method outperformed AlexNet, VGG16, ResNet50, MobileNetV2, and standard EfficientNet-B0, while ablation studies and Grad-CAM visualizations were used to assess the proposed attention-guided components.