The invention discloses an LDA (linear discriminant analysis) face recognition method based on nonlinear feature and model merging. The method comprises the following steps of (1) dividing a face data sample set; (2) extracting a face nonlinear feature set; (3) dividing a feature group; (4) building a face feature model base; (5) obtaining a global face feature model; (6) resolving a linear discriminant function to analyze projection matrix of the LDA; (7) projecting a face nonlinear feature set; (8) using a K Nearest neighbor classifier for performing classification recognition on the projected face feature. By using the non-linear feature extraction method, the more accurate face feature can be extracted; the higher recognition rate can be obtained; meanwhile, a method of building the model through feature grouping and obtaining a global model through merging is used; the matrix feature decomposition is avoided; the calculation time is short; the model can be repeatedly utilized; the method is more applicable to big data scene and distributive scene.