3D Facial Reconstruction Based on a Single Image Using CNN

In recent years, 3D facial reconstruction marked its presence in several areas, such as biometric security, computer vision, image synthesis, and video games. Many facial recognition applications require accurate 3D reconstruction; however, this task is complex and requires many calculations. This paper presents a new method based on CNN, which offers an automatic solution for the 3D reconstruction of the face depending on a single 2D image. Our method operates in three steps: the first step is to train a convolutional neural network with a self-made dataset to predict and detect the landmarks of the face and estimate the positions accurately from a single facial image in the image space. The second step involves producing the face’s geometric shape (mesh). Finally, the third step is to do an automatic translation between the 3D space of the object and the 2D image space to determine the texture referrals that correspond to each face polygon; thus, our method produces superior results in terms of accuracy and visual quality, considering all the parameters of the model (shape, expression, reflectance, and lighting) as inputs. Our method is simple, easy to implement, and offers a real-time 3D reconstruction of the face.