<p>Facial recognition (FR) has seen significant advancements due to deep learning (DL) techniques, which offer discriminative face representation and improved performance for various applications. However, traditional FR systems still struggle with lighting variations, facial expression changes, pose fluctuations, occlusions, and overfitting, which reduce accuracy and reliability. Illumination inconsistencies affect feature extraction, leading to misclassification, while facial expressions and pose variations distort facial landmarks, making recognition less effective. Occlusions, such as accessories or partial face coverage, obscure critical features, causing false detections. Additionally, many models suffer from overfitting due to limited training diversity, reducing their generalization capability. The automatic face recognition system proposed in this research uses transfer learning in unconstrained situations and is based on an enhanced VGGFace- 16 model. The three steps of the approach are data augmentation, face detection, and face recognition. The training dataset is more diverse and less prone to overfitting using data augmentation techniques such as random rotation, shifts, shear, resizing, contrast enhancement, rescaling, and flips. Robust face detection is achieved by applying the Histogram of Orientated Gradients (HOG) technique. A modified VGGFace- 16 model is used for face recognition to enhance performance, including layers such as Global Average Pooling and four thick layers with PReLu activation. The system's effectiveness is demonstrated through its high accuracy, precision, recall, and F1-score of 99.9% on the LFW and VGGFace2 datasets, thereby overcoming traditional FR problems.</p>