د.عبد المولئ ناجح2026-06-072026-06-07https://dspace.academy.edu.ly/handle/123456789/2173Face image matching for biometric authentication is a pivotal technology in digital security, yet it faces persistent challenges regarding cross-dataset generalization and computational efficiency in real-world environments. This thesis proposes an optimized deep-learning framework utilizing the EfficientNetV2-B0 architecture to establish a robust and lightweight face-matching system. Adopting a quantitative comparative experimental methodology, the study fine-tuned the model using the large-scale VGGFace2 dataset and rigorously evaluated its performance on the unconstrained Labeled Faces in the Wild (LFW) benchmark to assess generalization capabilities. The experimental results demonstrate the superiority of the proposed framework, achieving a Top-1 accuracy of 99.63% and a Top-5 accuracy of 99.94%, surpassing established state-of-the-art models such as Arc Face and ResNet-50. In terms of biometric security, the system recorded a False Acceptance Rate (FAR) of 0.004 and an Area Under the Curve (ROC-AUC) of 0.995, ensuring high reliability for identity verification. Furthermore, the model exhibited exceptional computational efficiency with an inference speed of 67.2 images per second and a compact size of 7.1 million parameters, making it 1.97 times faster than ResNet-50 and highly suitable for deployment on resource-constrained edge devices. This research confirms that optimizing lightweight convolutional neural networks effectively bridges the gap between high-accuracy academic prototypes and practical, real-time biometric applications.Face image matching for biometric authentication is a pivotal technology in digital security, yet it faces persistent challenges regarding cross-dataset generalization and computational efficiency in real-world environments. This thesis proposes an optimized deep-learning framework utilizing the EfficientNetV2-B0 architecture to establish a robust and lightweight face-matching system. Adopting a quantitative comparative experimental methodology, the study fine-tuned the model using the large-scale VGGFace2 dataset and rigorously evaluated its performance on the unconstrained Labeled Faces in the Wild (LFW) benchmark to assess generalization capabilities. The experimental results demonstrate the superiority of the proposed framework, achieving a Top-1 accuracy of 99.63% and a Top-5 accuracy of 99.94%, surpassing established state-of-the-art models such as Arc Face and ResNet-50. In terms of biometric security, the system recorded a False Acceptance Rate (FAR) of 0.004 and an Area Under the Curve (ROC-AUC) of 0.995, ensuring high reliability for identity verification. Furthermore, the model exhibited exceptional computational efficiency with an inference speed of 67.2 images per second and a compact size of 7.1 million parameters, making it 1.97 times faster than ResNet-50 and highly suitable for deployment on resource-constrained edge devices. This research confirms that optimizing lightweight convolutional neural networks effectively bridges the gap between high-accuracy academic prototypes and practical, real-time biometric applications.هندسة كهربائية