3D landmark-based face restoration for recognition using variational autoencoder and triplet loss

Abstract

Restoration of a 3D face from the mesh image is highly demanded in computer vision applications. 3D face restoration is a challenging task due to the variation of expression, poses, intrinsic geometries, and textures. The proposed technique consists of two main components, namely face restoration and recognition. A novel three-dimensional (3D) landmark-based face restoration method is proposed. 3D facial landmarks are used in the face recognition technique. It uses the principle of reflection and mid-face plane for the restoration of facial landmarks. By using the restored 3D face, a deep learning-based face recognition system is developed. It utilizes the concept of deep features from variational autoencoders. Further, these deep feature embeddings are trained using triplet loss training to increase the distance between embeddings of different persons and decreasing the distance between embeddings of the same person. These trained embeddings are used in support vector machine for prediction. The proposed framework is compared with recently developed face recognition techniques in terms of computational time. The proposed technique is able to recognize the person's face with better accuracy than the existing methods. Further, ablation studies are conducted to test the robustness of the proposed technique.