The Performance of Siamese Neural Network for Face Recognition using Different Activation Functions

The state-of-the-art method for image identification tends to achieve high performance. However, it requires a vast set of datasets making it infeasible to train machine learning models with sufficient samples while having newly-found samples. On the other hand, one-shot learning can be trained with limited samples. In addition to that, one-shot learning can also be trained with only one sample per person. One of the renowned methods is by using Siamese Neural Network. A Siamese network operates by having two identical networks with different images and learn the absolute difference between the two feature vectors while calculating the similarity score between the two images. However, the best activation function for this task is somehow unknown. Therefore, this paper attempts to evaluate the performance of Siamese Neural Network for face recognition using different activation functions. From the results, the most suitable activation function with the most stable performance is sigmoid, with an average accuracy of 92% for N-way Siamese Neural Network.