Exploring Face Recognition: A Comparative Study Using Loss-Based Methodologies

The field of face verification and recognition systems is rapidly evolving with significant advances being made. To clearly distinguish different face samples, the intra-class and inter-class similarity between samples must be clearly distin-guishable in latent space. The performance of a Deep Learning based face recognition system is primarily governed by the variety and sufficiency of data on which the model is trained and evaluated. The study presents and analysis on generalization capabilities of a Deep Learning model trained using a siamese network with different loss functions. We harnessed transfer learning through a pre-trained Resnet50 model, which acted as an embedding extractor. Our experimentation was conducted using the standard dataset “Labeled Faces in the Wild (LFW)” for both training and testing phases. We employed both offline and online triplet mining strategies within the framework of triplet loss-based training methodology. For contrastive loss-based methodology, the Euclidean distance serves as the metric to calculate the similarity between face image samples. In contrast, for triplet loss-based methodology, cosine similarity is utilized as the metric to ascertain similarity between embeddings. Distinct training instances were established, each implementing either contrastive loss or triplet loss as distinct loss functions. Our experimental findings distinctly favor the triplet loss-based model, surpassing the performance of the contrastive loss-based model and achieving an accuracy exceeding 70%.