<p>Accurate and discriminative feature extraction is essential for the effectiveness of biometric recognition systems. Finger Knuckle Print (FKP) recognition has emerged as a reliable biometric modality due to its distinctive line and texture patterns, stability, and suitability for contactless acquisition. The most advanced and effective methods for FKP recognition rely on convolutional neural networks (CNNs); however, conventional CNN architectures often fail to fully exploit inter-channel dependencies, which constrains their representational power. In this work, we enhance ResNet50V2 with lightweight Squeeze-and-Excitation (SE) blocks to strengthen channel-wise feature recalibration for FKP recognition. Four integration strategies—SE-Pre, SE-Post, SE-Identity, and Standard SE—are systematically evaluated by embedding SE modules at different stages of the residual architecture. This design allows a controlled assessment of how attention placement influences discriminative power and classification performance. Experimental evaluations on two benchmark datasets, IITD and PolyU v1, show that SE-ResNet50V2 consistently outperforms baseline networks, including AlexNet, MobileNet, and vanilla ResNet50V2. The proposed model achieved 98.73% accuracy on IITD (baseline: 98.10%) and 98.41% on PolyU v1 (baseline: 97.61%), demonstrating both improved accuracy and scalability across datasets. These results confirm that integrating SE blocks into ResNet50V2 provides a strong benchmark for biometric recognition, combining enhanced representational capacity. The framework thus offers a practical solution for real-world authentication applications requiring high precision and robustness.</p>