We use separable coded mask to replace the lenses, atop the image sensor-lensless imaging which can make system thinner. The imaging mathematic model is obtained by training a large number of separable patterns, and the original scene can be reconstructed by the computational imaging method, which can be used for the imaging of complex extended scenes under visible light. The method greatly reduces the thickness, volume, and weight of the imaging system. Currently, we build a prototype at visible wavelengths, successfully reconstruct images of 512 x 512 pixels in visible light. We will improve image quality and resolution to achieve the level of face recognition in future.