Back to index
Dynamic adaptation in an image
transcoding proxy for Mobile WWW browsing
Richard Han, Pravin Bhagwat, Richard LaMaire, Todd Mummert,
Veronique Perret, Jim Rubas, IBM Watson
A framework for determinig when/whether/how to transcode in a
TranSend-like proxy is presented.
Overview/Main Points
- Output size of an image in bytes is a linear function of
input size in pixels with correlation >0.9.
- JPEG xcoding is DCT-dominated and latency is correlated linearly
to image area in pixels, with rho >0.98
- A bits per pixel value of around 0.1 is representative of
"cartoon like" GIFs and logos.
Relevance/Observations
- Most images are now small enough that the other overheads seem
likely to dominate (like connection establishment), so
repackaging (TCTM-style) should work better.
- Authors try to pick a parameter vector for xcoding that causes
the xcoding step to finish within a preset time limit. This
turns out to be quite hard to do. Instead, we
could do "progressive distillation" till we run out of time,
i.e. do multiple passes. This is probably not worth it for
frequency-domain or layered encodings,
but may be worth it when the various
"passes" do different things, e.g. a pass that shrinks the image
followed by an optional pass that extracts text using OCR or something.
Back to index