RF04-07 Analysis of Deep Embeddings for Facial Recognition Using Vector Databases and Synthetic Data Generation Details

Start Date: 2026-03-10

End Date: 2026-03-10

Deep learning–based facial recognition has experienced significant advances in recent years due to the development of models capable of generating highly discriminative vector representations of human faces. These representations, known as deep embeddings, allow identities to be compared through similarity metrics within high-dimensional vector spaces. Modern facial recognition frameworks such as DeepFace generate embeddings from facial images. These embeddings can later be stored and searched using vector similarity engines such as FAISS, enabling efficient similarity search among large numbers of facial representations. This project proposes the development of an experimental prototype to analyze the geometric structure of facial embedding spaces. The main objective is to evaluate how synthetic data generation influences the distribution of vectors and the separability between identities within the representation space. The methodology involves extracting embeddings from facial images, generating synthetic samples through image transformations, storing embeddings in a vector database, and analyzing intra-class and inter-class distances between embeddings. The expected results aim to demonstrate that synthetic data improves identity separability and increases the robustness of facial recognition systems. Additionally, the project evaluates the use of vector databases as efficient infrastructures for storing and querying high-dimensional representations in artificial intelligence applications.