In this project, a lightweight light estimation strategy based on penumbra effect is proposed, which optimizes the DeepLight: light source estimation for augmented reality using deep learning 's scheme. Because the original method only estimates the direction of a single main light source, the effect is not ideal. Considering that the soft shadow effect is formed by multiple light sources in different directions. In this paper, the estimation of single main light source in the original method is improved to the estimation of multiple parallel light directions, and the concept of color temperature of light source is added. The model is trained using batch rendering RGB-D synthesis data randomly generated by Blender, in which the image data uses a channel (transparent channel) to store the distance between the object and the camera (depth information). In order to make the composite data closer to the real image, a number of post-processing operations are carried out on the image after rendering.