Abstract: Two-phase test sample representation algorithm (TPTSR), which is robust to interference such as occlusion and noise, performs well in face recognition without pose variation. However, its recognition rate will decline when the face pose varies dramatically. To solve this problem, a three-phase test sample representation algorithm was proposed. The first was frontal face synthesizing was a frontal face with small horizontal deflection angle was synthesized using view-library and proposed frontal face synthesizing algorithm. Thus, a frontal face was synthesized as the new test sample. The second was training sample selecting phase, M training samples that make the most contribution were selected to represent the new test sample. The third was decision and recognition phase, a face was recognized using the M training samples. Experiments on some publicly available face recognition benchmarks demonstrate that the proposed 3PTSR algorithm outperforms the state-of-the-art methods in challenging conditions, especially for the face with various poses.