The invention provides a three-dimensional face recognition method based on semantic alignment multi-region template fusion, which comprises the following steps of: 1, determining a data set of registered faces and test faces in a three-dimensional face database; 2, preprocessing all registered and to-be-identified three-dimensional face models, and performing dense alignment on the preprocessed three-dimensional face models and the reference model; 3, pre-dividing the face region into a plurality of template regions which do not contain expression influence and can be overlapped; 4, for eachtemplate area, directly calculating a similarity value between the template areas on the three-dimensional structure of the human face; And 5, independently voting each region according to the similarity value, synthesizing a plurality of region matching results, and determining a final matching result by adopting a majority voting mode. According to the face recognition method provided by the invention, similarity prediction is carried out by utilizing mutual independence of the multi-template areas, the dependence of an algorithm on accurate division of a single area is reduced, and meanwhile, a multi-area template common voting strategy is adopted, so that certain robustness is also achieved on expressions and other factors influenced by the areas.