Pose, Expression, Illumination Invariant 3D Face Recognition Based on Transfer Learning

Since the last decade, 3D face recognition is much more popular than 2D face recognition. The face recognition problem related to pose variation is easily solved by 3D face recognition. In this work, we have introduced a 3D face recognition system using a transfer learning-based approach. Deep learning is one of the most popular techniques in classification or recognition. The convolution neural network (CNN) in deep learning is mainly used for image-based recognition. However, a heavy dataset and high configuration machine are basic needs to develop a deep learning-based system. Due to the limitations of 3D databases consisting of a vast number of data, we have focused on a transfer learning-based approach on five different pre-trained network models such as ResNet50V2, InceptionV3, InceptionResNetV2, DenseNet201, and Xception. This approach transfers knowledge from a similar task to a new task for improving performance. Before feature extraction using the transfer learning approach, a binarization-based cropping technique was applied to a 3D depth image to remove outliers and extract the cropped face portion. The proposed 3D face recognition system considered two popular 3D face databases, Frav3 and Casia3D, for experimental analysis. The accuracy, recall, precision, and F1-score are calculated for all the pre-trained models for two individual databases.