Digital human expression modeling method based on geometric deep learning

The invention relates to a geometric deep learning-based digital human expression modeling method. The method comprises the following steps of 1, collecting a three-dimensional human face image set and performing normalization processing on images in the three-dimensional human face image set; 2, extracting geometric deformation features G of a face region from each three-dimensional face image, and representing a deformed face Mdef through the geometric deformation features G; 3, inputting the geometric deformation feature G into an identity and expression decomposition model to obtain an identity geometric feature G'id and an expression geometric feature G 'exp; and step 4, recombining the identity geometric feature G'id and the expression geometric feature G 'exp through an aggregation network to complete digital human expression modeling. The method has the advantages that a three-dimensional image can be used as continuous input, identity features and expression features representing a human face are sep