Transformer-based Siamese Neural Network Face Verification Method

Abstract

Face verification, as a key part of person intelligence tracing, aims to determine whether the same person is in two face photos, which can be approximately transformed into a binary classification problem. In this paper, the siamese neural network architecture is adopted as the overall structure of the method, and the traditional bacbone network generally uses a convolutional neural network, this paper will use Vision Transformer as the backbone network as the method. Compared with convolutional neural network, its multi-headed attention mechanism can better focus on the larger sensory field and extract face features, which can then better perform face verification, while Binary cross-entropy is used as the loss function. After the validation of the private dataset, the model can recognize faces well and has excellent validation effect, which is of great value for the field of human intelligence tracing.