Stray light (SL) control is an important aspect in the development of optical instruments. Iterations are necessary between design and analysis phases, where ray tracing simulations are performed for performance prediction. This process involves trial and error, requiring to be able to perform rapid evaluation of SL properties. The limitation is that accurate SL simulations require sending many rays, which can be time consuming. In this paper, we use deep learning to improve the accuracy of SL maps even when obtained with very few rays. Two different deep learning methods are used.The training process is performed by generating a large database of artificial SL maps, with different noise levels reproduced with a Poisson distribution. Once the training completed, we show that the autoencoder performs the best and improves significantly the accuracy of SL maps. Even with extremely small number of rays, it recovers complex SL patterns which are not visible on raw ray traced maps. This method thus enables more efficient iterations between design and analysis. It is also useful for developing SL correction algorithms, as it requires tracing SL maps under large number of illumination conditions in a reasonable amount of time.