The presentation of a semi-supervised deep learning platform for 3D face reconstruction from 2D images

In recent years, 3D face reconstruction approaches based on deep learning have had good results, which perform well in terms of quality and efficiency. In this paper, we present a semi-supervised deep learning platform for 3D reconstruction from 2D images, where we use two pre-trained unsupervised segments to reduce the need for 3D and 2D labels. In this way, by using the trained parts for the training of the entire network, we need less labeled data. The original goal of the presented platform is to find a map between two-dimensional representation spaces and three-dimensional representation spaces with lower dimensions. Therefore, the proposed platform in this article includes the unsupervised parts of the mapping from 2D and 3D spaces to the low-dimensional representations and the supervised part of the mapping between the representations with low-dimensional. Generally, therefore, we present a method that: 1) uses a stout and combined loss function for weakly supervised learning that considers the low-level information and the data at the perceptual level for learning; 2) applies the multi-image face renovation by using the supplementary data from the different pictures to consolidate the form. In order to prove the effectiveness of the proposed method, the various tests have been performed on three data sets. The results of these tests show that the proposed method significantly reduces the amount of the reconstruction error in compared to the similar methods.