A Novel Face Recognition Algorithm for Imbalanced Small Samples

Abstract

Deep learning (DL) has become a hotspot in the research of image recognition. However, the DL strategy must be trained with lots of samples that are distributed evenly across classes, i.e. subjected to balanced distribution. Therefore, this paper attempts to design a method to satisfactorily recognize faces in imbalanced small samples. Firstly, the deep convolutional generative adversarial network (DCGAN) was improved to generate data samples with similar distribution as the original training data, creating a balanced training set of sufficient labelled samples. Then, transfer learning was performed to transform the AlexNet, which is pretrained on big dataset, to the balanced target dataset of small samples. Next, the previous convolutional layer was frozen as a feature extractor, and the truncated normal distribution was reinitialized for the next fully-connected layer. Simulations on face recognition show that our method achieved higher recognition rate and less serious overfitting than ordinary CNNs.