The invention discloses a vehicle model identification method based on reinforcement learning, and aims to fully excavate the advantages of CNN in the aspects of feature extraction and classificationidentification, and provide a solution for realizing accurate and robust vehicle model fine identification. The vehicle model identification method is characterized in that a human visual attention mechanism is simulated; a visual attention model based on strong classification significance and a viewpoint automatic selection method are designed; and the visual attention model is characterized in that an attention mapping matrix and a viewpoint focusing template are constructed, and the viewpoint automatic selection model is characterized in that a reinforcement learning algorithm based on SARSA is used for autonomous selection of viewpoints, so that the model can adaptively select an optimal image recognition area, and an optimal vehicle model recognition effect is obtained. The vehicle model identification method not only can overcome the defect that a traditional manual feature extraction algorithm cannot adapt to changes of the position, the scale and the contour of the vehicle in the image, but also can effectively cope with challenges brought by camera shooting angle changes and vehicle shielding.