Research on 3D reconstruction based on 2D face images.

Abstract

3D face reconstruction is a popular research area in the field of computer vision and has a wide range of applications in various fields such as animation design, virtual reality, medical guidelines, and face recognition. Current commercial 3D face reconstruction generally relies on large image scanning equipment to fuse multiple images through sensors for 3D face reconstruction. However, this approach requires manual modelling, which is costly in terms of time and money, and expensive in terms of equipment, making it unpopular in practical applications. Compared to 3D face construction with multiple images, the single-image approach reduces computational time and economic costs, is relatively simple to implement and does not require specific Hardware equipment. Therefore, we focus on single-image approach in this dissertation and contribute in terms of research novelty and practical use. The main work is as follows： A unique pre-processing process is designed to separate face alignment from face reconstruction. In this dissertation, the Active Shape Model (ASM) algorithm is used for face alignment to detect the face feature points in the image. The face data is posing corrected so that the corrected face is better adapted to the face pose of the UV-Position Map. The UV coordinates are then used to map the 3D information onto the 2D image, creating a UV-3D mapping map. In order to enhance the effect, this dissertation also does face cropping to fill the whole space as much as possible with face data and expands the face dataset using rotation, scaling, panning and noise addition. Improving the neural network model by using the idea of residual learning to train the network model incrementally, emphasizing the reconstruction of the model for deep information. Face data characteristics are first extracted using the encoding and decoding layers, and then face features are learned using the residual learning layer. By comparing with the previous algorithm, we achieved a considerable lead on the 300W-LP face dataset, with a 35% reduction in NME error accumulation over the RPN algorithm. Based on the pre-processing methods and residual structures we proposed, the experimental results have shown good performance on 3D reconstruction of faces. The end-to-end approach based on deep learning achieves better reconstruction quality and accuracy compared to traditional, model-based face reconstruction methods.