Deep learning-based multiview face three-dimensional model reconstruction method

Abstract

The invention discloses a deep learning-based multiview face three-dimensional model reconstruction method, and belongs to the field of computer vision. The method comprises the following steps of: generating a multi-illumination multiview virtual face image; generating a depth map of a face front view; training a plurality of independent and parallel convolutional neural networks; training a neural network in which weights of various views are distributed; and restoring depth maps output by the networks into a face three-dimensional grid model and carrying out peak coloring. According to themethod, multiview images are independently trained to restore depth maps, and each view weight distribution map is trained to carry out deep integration, so that the face three-dimensional model reconstruction precision is improved under the premise of ensuring the efficiency.

Description

technical field [0001] The present invention relates to the field of computer vision, in particular to a method for reconstructing a multi-view human face three-dimensional model based on deep learning. Background technique [0002] The 3D face model is widely used in security authentication, film and television animation, medical science and other fields. However, the cost of obtaining real and detailed facial information is very expensive, such as using a 3D laser scanner. However, using multi-view to reconstruct the 3D face model based on deep learning has the advantages of high speed and low cost. The proposed image-based 3D face reconstruction algorithms can be roughly divided into two categories: [0003] The first type of method is based on multi-viewpoint 3D face reconstruction. Usually, the traditional method is used to obtain the multi-view image of the face first, then use the structure from motion algorithm (Structure From Motion) to calibrate the camera param...

AI Technical Summary

Problems solved by technology

[0006] The above-mentioned prior art has at least the following disadvantages: the selected training data of almost all methods based on deep learning is based on the three-dimensional deformation model proposed by Blanz and Vetter, and usually the 3DMM face parameters are used as the input of the neural network for network training. Predict the 3DMM face parameters corresponding to the input image and restore the 3D model of the face

Since 3DMM uses the principal component analysis (PCA) method to build a statistical model, and PCA is essentially a low-pass filter, the effect of this type of method is still not ideal in restoring the detailed features of the face.

However, traditional methods tend to be slower in processing speed, or need to input additional information

Images