Multi-feature output deep network ensemble learning for face recognition and verification

Abstract In facial recognition systems, the sample size of face data is often limited, making it difficult to improve recognition performance through transfer learning using pre-trained deep neural network models (DCNNs). Furthermore, in complex environments, the recognition performance of a single DCNN is greatly weakened. To address these issues, this paper proposes a multi-feature DCNN ensemble learning method based on machine learning for facial recognition and verification. First, the characterization ability of face details is enhanced by expanding the output feature dimension of the DCNN model. Then, machine learning methods with few or no parameters are used for secondary learning on small sample databases to improve recognition/verification performance.In this work, we utilized several mainstream network models such as ResNet50 and SENet to expand and integrate features. The experimental results show that our proposed method can improve the recognition/validation accuracy of existing DCNN models by up to 10%. Moreover, this method does not require training the network, has relatively low computational costs, and has a wide range of practical applications.