Figure 4: Using averaging kernels for downscaling: (b) linear filters
associate the same averaging kernel with every output pixel,
forcing a global compromise between smoothing and aliasing. (c)
unoptimized bilateral kernels adapt to the image, but might “miss”
image features depending on the placement of their centers (e.g.,
the dial plate numbers). (d) Our optimized kernels adapt to the
image even more, but do not suffer from missed or disconnected
features.