Augmented reality (AR) has been widely applied in medical education that allows for digitally generated three-dimensional representations to be integrated with physical training module. However, correcting color distortion in projected displays is a major challenge. In this paper, we propose a convolutional neural network based projector photometric compensation framework to compensate for color shift. The architecture is a generative network cascading a decoder subnet. Such architecture compared the color shift of the projection from the input image, as well as the original textures. We used a custom projector with low-illumination for our experiments. To compensate content losses of the projection, we utilized a sharpening module at the end of network output. Additionally, to adapt the target projection screen using small dataset, we synthesize solid color images automatically for network training. The experiments results show that our solution of projector photometric compensation outperforms state-of-the-art method.