Modified spectral clustering algorithm for semisupervised face annotation modeling

Recent advancements in facial recognition have impacted security, healthcare, and identity verification, with face annotation — labeling facial features for training datasets - being crucial. However, traditional annotation is time-consuming and labor-intensive, making sufficient accurately labeled data hard to obtain. This research addresses this via semi-supervised learning, using semi-supervised clustering to expand datasets with limited labeled samples. A key issue in spectral clustering is fixed similarity matrices, which struggle with high-dimensional facial data complexities, non-linear relationships, and noise, leading to suboptimal results. To improve accuracy, an optimized similarity matrix was developed. It also integrated MustLink (ML) and Cannot-Link (CL) constraints, generated efficiently via the Label Propagation Algorithm (LPA) enhanced by combining Share Nearest Neighbour (SNN) and Radial Basis Function (RBF). Objectives included enhancing LPA accuracy, developing robust constraintbased clustering, and evaluating performance. The research streamlines annotation, reduces manual work, boosts facial recognition performance (98.97% purity), and contributes to computer vision and AI with efficient large-scale face annotation solutions, developing two models.