In this study, we propose an inverse reflectance model based on co-located images to precisely model the nonlinear reflection behavior of the non-diffuse reflective surfaces. The proposed model can accurately map the pixel value to the product of the normal vector and the light direction. We need to capture only one co-located image and one RGB image under multispectral conditions to ensure that photometric stereo vision can achieve a high-precision performance, so the time required to capture images is considerably reduced. To perform surface inspection in case of mass production, the proposed method can realize online detection of the moving surfaces at a microsecond shooting rate because the co-located image can be acquired in advance and used for the subsequent workpiece. However, the iterative steps applied in the traditional methods are omitted, and the robustness with respect to outliers, such as shadow points and highlights, is improved, because a neural network is used in the proposed method to train the near-field photometric stereo model. Furthermore, the results of simulation and experiment show that the algorithm can recover the normal vector of the non-diffuse surface well under the condition of very few images.