A novel LDA method for Tongue Images Recognition (TIR) based on a combination of global features extracted by nonlinear kernel principal component analysis and local features derived by applying Gabor wavelets is discussed. It is well known that the distribution of tongue images is highly nonlinear under a large variation in viewpoints. Therefore, linear methods such as Principle Component Analysis or Linear Discriminant Analysis cannot provide reliable and robust solutions for TIR problems. In our framework, the improved LDA in the unitary space makes use of the null space of the withinclass scatter matrix effectively, and Global feature vectors and local feature vectors are integrated by complex vectors as input feature of the improved LDA. The experiment results demonstrate that the proposed methodology is more effective and robust for tongue images recognition.