Sparse Representation-Based Face Object Generative via Deep Adversarial Network

Abstract

How to generate well quality faces objects of automated processes has always been the focus on researchers. Recently, due to the deep generative networks have achieved impressive successes in data generative fields, researchers have tried to introduce deep learning into the 3d objects generate field, such as text2scene, slice-based object generate. However, the generative ability in 3D object is limited by the size of the feature space, because of computational space limitations on hardware. In this paper, we address the problem by reducing amount of calculated on process of learning, and thus generative newly different objects. The problem is intractable, since first the limiting of compute space is so hard that object can't be process in deep network due to the process need to compute many matrix multiplications. To resolve the problem, we propose a sparse representation-based method of generating well-quality faces object. Our method consists of two parts: sparse reconstruction and object generative. First, we verified the possibility of using sparse representations of 3D data by reconstructing 3D object. Second, we design a network architecture of deep adversarial network of generating new sparse representation and combined with the previous reconstruction method of generating new face object. Experiments show that our method has the ability to generate very different and well quality faces objects that contain tens of thousands of points and meshes. Our findings show that sparse representation can be used in 3D object reconstruction and generate via deep generative adversarial model.