Influencing Factors in Facial Recognition and Verification: A Siamese Neural Network Hyperparameter Study

There are many practical uses for facial verification and recognition jobs. In this study, we look into the developments and difficulties faced in facial recognition technologies, especially in the application of Siamese neural networks. This study converges around optimizing hyperparameters to improve Siamese network accuracy and performance. We evaluated the impact of different optimizers, loss functions, and data preprocessing techniques on Siamese network performance through a chain of experiments. The results demonstrated the importance of choosing the right hyperparameters, demonstrating how decisions such as the Adam optimizer and Binary Cross-Entropy as the loss function may greatly enhance the system's capability to precisely identify and validate faces. This study also overlooks the importance of diverse and high-quality datasets and states how these elements affect the model's performance. With this work, we hope to give a thorough manual for fine-tuning Siamese neural networks for use in facial recognition applications, along with helpful advice on how to overcome the difficulties posed by dataset variability and hyperparameter selection. In addition to adding to the theoretical knowledge of facial verification technologies, our research offers useful suggestions for enhancing the conception and application of these systems in actual settings.

I. Introduction

In this era, where technology and Artificial Intelligence (AI) have reached unprecedented popularity, deep learning [1] emerges as an important subject of interest. As a branch of machine learning, deep learning's acclaim stems from its remarkable capability to learn from and make predictions based on vast amounts of data. Currently, the domain of face verification [2] and recognition is capturing the attention of researchers and practitioners alike, motivated by the potential to apply AI in addressing varied real-world challenges [3]–[5], e.g., in medical applications [6], financial applications [7], linguistics [8], [9], poverty estimation [10] and others [11]–[13]. Techniques such as Convolutional Neural Networks [14]–[16] and Siamese Neural Networks [17] are at the forefront of advancements in facial recognition technology, showcasing impressive outcomes in this rapidly evolving field. Siamese neural networks are a type of deep learning architecture that has excelled at verifying and identifying facial features. Siamese networks are created to learn a similarity metric between two images, as opposed to traditional neural networks, which are trained to classify images [17] into predefined categories. They are therefore perfect for facial verification, which is the process of comparing two images to see if they belong to the same person. Siamese networks are superior to conventional networks in their ability to handle variations in lighting, pose, and facial expressions and their capacity to generalize to new faces with little training data. Table I displays the summary of the literature for facial recognition.