부산대학교 도서관

Fast and robust 3D reconstruction of facial geometric structure from a single image is a challenging task with numerous applications, but there exist two problems when applied "in the wild": the 3D estimates are unstable for different photos of the same subject; the 3D estimates are over-regularized and generic. In response, a robust method for regressing discriminative 3D morphable face models(3DMM) is described to support face recognition and 3D mask printing. Combining the local data sets with the public data sets, improving the exiting 3DMM fitting method and then using a convolutional neural network(CNN) to improve reconstruction effect. The ground truth 3D faces of the CNN are the pooled 3DMM parameters extracted from the photos of the same subject. Using CNN to regress 3DMM shape and texture parameters directly from an input photo and offering a method for generating huge numbers of labeled examples. There are two key points of the paper: one is the training data generation for the model training; the other is the training of 3D reconstruction model. Experimental results and analysis show that this method costs much less time than traditional methods of 3D face modeling, and it is improved for different races on photos with any angles than the existing methods based on deep learning, and the system has better robustness.