We have presented a novel content-adaptive image downscaling
method that adapts the shape of its downsampling kernel, yielding
sharper and more detailed downscaled results. Contrary to common
wisdom that dictates that frequencies above the Nyquist frequency
introduce artifacts in the downsampled image (in the form of aliasing),
we show that by careful sampling, certain high frequencies
features can still be preserved in the downscaled image without artifacts.
Given the growing “resolution gap” between cameras and display
devices and the advent of gigapixel panoramic imaging, we believe
that this work opens up an exciting area of research. There are
plentiful avenues for future research. Our work has shown that it
is possible to sometimes drastically improve quality over existing
downscaling methods. For future work we would like to further improve
the robustness of the method, e.g., through smarter heuristics,
so that our method always outperforms simpler filters. It would also
be interesting to look at other signals than images as inputs. A natural
immediate step would be to analyze and constrain the temporal
behavior of our algorithm, e.g., when applying it to videos.