Shallow focus can emphasize one part of the image over another. Our trained convolutional neural network estimates dense defocus map that represents the size of circle of confusion (CoC). By observing the small and large context simultaneously, the network rarely misses the smooth but focused area.