Deep Convolutional Neural Networks with Transfer Learning for Neonatal Pain Expression Recognition

For the neonatal pain expression recognition task, the recognition precision of the algorithms based on traditional machine learning isn't robust to the illumination and pose variations. The recognition algorithms based on deep learning usually rely on large-scale labeled training datasets, the recognition performance of these algorithms will be low when the labeled neonatal pain expression image dataset is small. To overcome these drawbacks, we present a neonatal pain expression recognition approach based on pre-trained deep convolutional neural network (DCNN) model with transfer learning. In this work, the introduction of transfer learning technology avoids the occurrence of over-fitting and accelerate the training procedure. Firstly, some typical DCNNs which have been trained on the ImageNet dataset, such as AlexNet, VGG-16, Inception- V3,ResNet-50 and Xception, are selected as the basic models to extract the general features of images. Then, in order to enhance the generalization ability of the DCNNs, the pre-trained DCNNs are fine-tuned by using the neonatal pain expression image dataset, and so that the feature transfer from the general image to the neonatal expression image is realized. Finally, we use different transfer learning methods to test the fine-tuned DCNN models. The experiment results show that the fine-tuned VGG-16 model achieved the best recognition accuracy (78.3 %) on the small neonatal pain expression image dataset, which indicates that the fine-tuning method can effectively obtain a DCNN model with good performance, and the transfer learning is an effective method for training DCNN when the available labeled training dataset is small. The effectiveness of DCNN and transfer learning for neonatal pain expression recognition shows promising application for clinical diagnosis.