Fast Space-Variant Defocus Point Spread Function Estimation from a Single Image

Defocus blur is a common phenomenon in existing optical systems. However, due to the mutual coupling of light intensity information at each pixel position, it is difficult for existing methods to quickly and accurately calculate the defocus amount at each pixel position. In this paper, we proposes a model based on multi-scale supervision to quickly predict the defocus map with spatially varying characteristics, that is, to predict the point spread function of each pixel position. By combining synthetic data that is closer to the actual situation for training, the model shows the best performance at a prediction speed of 39 frames with a mean square error (MSE) and mean absolute error (MAE) of 0.00051 and 0.00858, respectively.