Enhancing Recognition of Blurred Faces through FaceNet and Attention Mechanisms

Facial recognition is a rapidly evolving field with applications ranging from smartphone authentication to public safety and surveillance. While deep learning has significantly improved recognition performance under ideal conditions, real-world deployments often fall short—especially when processing blurry images that reflect real-world conditions. These images, frequently captured from surveillance footage, are affected by motion blur, poor lighting, low resolution, and occlusions. Such distortions compromise facial feature clarity, making it difficult for conventional models to extract reliable identity cues, ultimately reducing recognition accuracy. To address this, we propose a hybrid approach that enhances recognition accuracy by integrating Generative Facial Prior GAN (GFPGAN) for image restoration with transfer learning from the InceptionResNetV1 architecture of FaceNet. We also introduce a tailored attention mechanism and a progressive unfreezing training strategy. The attention mechanism effectively combines spatial and channel-wise components, enabling the model to focus dynamically on the most informative facial regions—such as the eyes, nose, and mouth—while downplaying less relevant background noise. This targeted emphasis boosts feature extraction, especially under degraded visual conditions. Our progressive unfreezing strategy further strengthens the model by gradually unfreezing deeper network layers during training, allowing the model to retain high-level pre-trained knowledge while adapting incrementally to the specific challenges of blurry surveillance data. This technique supports stable learning and reduces the risk of overfitting, which is particularly important when working with smaller or noisy datasets. We evaluate our method using a curated Surveillance dataset composed of blurry facial images taken under real-world conditions. The proposed model achieves a validation accuracy of 87.13%, marking a substantial improvement over traditional face recognition approaches in similar settings.