CS631 Paper Response Form
Paper Title: Algorithms for Manipulating Compressed Images
Author(s): Brain C. Smith and Lawrence A. Rowe
Main Point(s):
Linear image transformations can be performed on jpeged data
without fully decompressing. More specifically, only the
last step of the jpeg algorithm, namely entropy encoding of the runlength
encoded DCT coefficients, has to be undone. The application of
scalar multiplication, scalar addition and pixelwise addition to
the RLE vectors is quite straight-forward. Nice side-effects of using
the undecompressed data are that in the scalar multiplication, only
the non-zero values are actually multiplied, and that scalara addition
only affects the first DCT coefficient. Pixelwise multiplication
is more complicated but can be made fast by precomputing a sparse
representation of a matrix that depends on the quantization tables
that are used for the two input images and the output image. This technique is shown to yield a speedup of 100 for a dissolve operation and a
speedup of 50 for a subtitle operation.
Possible Use(s):
This technique should make editing of high resolution images possible even on machines
with few memory, low bus-bandwidth or a slow processor: the user designs
an effect on decompressed thumbnail-versions of the data and after
he or she is content with the prototype, the operation on the
full-sized data is performed on the compressed data.
Extensions:
A similar technique for MPEGed data would be the most desirable extension,
because this technique is obviously good for processing video.
Other extensions would be to build
higher level functions from the low-level linear operations and
optimize them, for example by avoiding the quantization-table
scaling in the intermediate steps. It would be very interesting to
develop a similar technique for non-linear image processing.
CS631 home page