What I believe is being done is interpolation in the frequency domain to increase the number of pixels while smoothing the transition from black to white.
Consider a single row for example, which is a rectangular function in time. The Fourier Transform would be a Sinc function, which would dominate the outer part of the spectrum in the frequency domain (go toward zero in the middle, which represents $F_s/2$ where $F_s$ is the sampling rate, and then rise back as a mirror of the first half up to $F_s$. Adding zeros in time simply interpolates more samples of this same function in frequency. Then truncating the center of the spectrum reduces the higher frequency range of the Fourier Transform, which effectively makes the pulse wider (lower frequency).
I discuss zero padding for FFT's in more detail here: What happens when N increases in N-point DFT