Face recognition in real-world scenarios poses challenges due to diverse image qualities and varying noise levels. To improve the Deep Joint Distribution Optimal Transportation (DeepJDOT) model's performance, we propose a two-step preprocessing strategy involving facial detection and cropping. Our goal is to achieve highly precise face recognition in the presence of different noise environments. We curate a dataset of 18 Hollywood celebrities with diverse appearances for evaluation. To enhance the DeepJDOT model, we adopt a sophisticated neural network architecture, enabling it to capture intricate facial features. Leveraging optimal transport, the model aligns class representations between domains and adapts to various noise scenarios. Through rigorous evaluation, our enhanced DeepJDOT model demonstrates superior face recognition performance, setting a new standard for robustness in real-world scenarios.