The curved surface projection model in fisheye image correction algorithm is presented. To analyze the causes of distortion in existing models, we establish an ideal surface projection model and compare its surface with the surfaces of existing models. Subsequently, feature points are obtained on the ideal surface according to the relationship of coordinates of fish-eye image points and their ideal three-dimentional (3D) points. Finally, the least square method is used to obtain a quadric surface and presents a quadric surface projection model. The experiment shows that the corrected image of the new model is more similar to the actual scene than the corrected images of previous models.