Figure 5
From: Progressive compressive sensing of large images with multiscale deep learning reconstruction
Illustration of the multiscale property of the 2D Hadamard transform (left) and the image (right). The upper left quadrant of the 2D Hadamard transform of an image (in blue) is a Hadamard transform of a lower resolution version of the image. Therefore, we can take the Hadamard samples highlighted in blue and reconstruct from them the 512 by 512 image. In turn, the upper left quadrant of the quadrant (in orange) is also a Hadamard transform of an even lower resolution image. Now, if we take the Hadamard samples highlighted in orange (or a quarter of the samples highlighted in blue) we can reconstruct the 256 by 256 image, directly from the Hadamard samples.
