Kernel Modeling Super-Resolution on Real Low-Resolution Images

Ruofan Zhou     Sabine Süsstrunk
Image and Visual Representation Lab
École Polytechnique Fédérale de Lausanne

- Supplementary Material -

project page

On this page we visualize the blur kernels (with size of 25x25) that we use in the paper. All kernel images are normalized by subtracting the minimum value of the kernel and then dividing by the maximum value of the kernel. The images will zoom when the mouse is on top.

Here are the bicubic downsampling (x2) kernel (implemented by Matlab), Gaussian kernel with σ=1.25, Gaussian kernel with σ=1.6 and Gaussian kernel with σ=1.7 (mentioned in Section 4.2 and 4.3 in the paper):

Note as the bicubic kernel contains negative value around the center, the background of its visualization is not black because of the normalization.

And below is the visaulization of the bicubic downsampling (x4) kernel and Gaussian kernels with σ=2.3, 2.5 and 2.7:

Bellow we show 20 examples of the estimated kernels from the bicubic interpolated (X2) low-resolution images (Section 3.2.1 in the paper):

Bellow we show 20 examples of the generated kernels (X2) from the GAN (Section 3.2.2 in the paper):

Bellow we show 20 examples of the estimated kernels from the bicubic interpolated (X4) low-resolution images:

Bellow we show 20 examples of the generated kernels (X4) from the GAN: