At present, the efficiency and accuracy of face recognition based on visible spectra are strongly influenced by the factors such as lighting, pose, occlusion, expression change and photo fraud, especially the face recognition problems in the night vision environments need to be solved. A face feature extraction algorithm based on thermal infrared spectra is proposed. The thermal infrared face images are modeled to obtain a maximal data model, and the mixed model parameters are estimated and adjusted as a Gaussian mixture model. The isothermal features of thermal infrared Gaussian mixture face images are extracted, and the thermal feature images are reconstructed. The similarity between individuals measured by the calculation of probabilistic proximity index is used to present the recognition results. The experimental results based on the UCHThermalFace database show that the proposed method has relatively high recognition precision and robustness to multiple attitudes, feature changes, random occlusion and eye noise samples, which greatly improves the anti-interference ability of face recognition system in night vision environments.