The idea of subpixel convolution was originally proposed by W. Shi et al from Magic Pony, and well explained in this post. The figure below illustrates in 1-dimensional case that subpixel convolution is probably just transposed convolution (or deconvolution).
