Efficient Transfer Learning Combined Skip-Connected Structure for Masked Face Poses Classification

Abstract

Aiming at the new requirements of masked face poses classification during the epidemic outbreak, this paper proposes an efficient transfer learning approach combined skip-connected structure to improve the accuracy of masked face poses classification in the absence of masked face poses data. We have worked on the following two aspects: 1) According to the features transition of the convolutional neural networks, we propose an efficient transfer learning approach and opt for a more appropriate source domain to solve the problem that the specificity of features in the pre-trained deep networks will damage the performance when transferring to the target domain. First, a semisynthetic masked face poses dataset is constructed to replace ImageNet as the source domain, which can reduce the span of transfer and improve the pertinence of transfer learning. Second, the shallow networks which contain the general features are frozen while the deep networks which contain the specific features are retrained and the entire networks are fine-tuned afterwards. It optimizes the specific features in the source domain when transferring, and promoted transfer learning more effectively; 2) To further improve the overall accuracy by improving the accuracy of masked face pose classes with subtle differences, a skip-connected structure is proposed to fuse general features containing rich detailed information in the shallow networks into the classifier. Experiments on AlexNet and VGG16 show that the proposed method has certain advantages, and the overall accuracy can reach 96.43% and 99.29% at the final respectively.