SNResNet: A New Architecture Based on SqNxt Blocks and Rish Activation for Efficient Face Recognition

Abstract

In this paper, we present a novel face recognition architecture based on the Inception-ResNet framework, called SNResNet. The Inception-ResNet architecture, is effective in computer vision applications but exhibits limitations such as computational complexity, high memory consumption, and data dependency. It uses the ReLU activation function and softmax loss function which are not best-suited for face recognition. The proposed SNResNet uses triplet loss as the loss function to be able to train the model on large datasets. The advantages of the triplet loss over the softmax are handling one-shot learning, robustness to class imbalance and fine-grained discrimination. The ReLU activation function rejects all negative values that in some applications reduce the accuracy of the model. To overcome this problem, we introduced a new activation function called Rish which has better performance. In addition, we optimized the Inception-ResNet-B block using the SqNxt block to control the model's computational costs. The CASIA-WebFace dataset is used to train the models. This dataset has some challenges; e.g., some photos have more than one face, and all faces have a background. Preprocessing conditions are defined to identify and align the correct face. SNResNet achieves 94.63% accuracy on CASIA-WebFace. Performance evaluation on the LFW benchmark database yields an impressive accuracy of 99.68%, surpassing the standard model's accuracy of 98.85%. Further, we reduced the FLOPS of the Inception-ResNet model by 15.61% which indicates a lower computational cost and a faster model for face recognition.