[magick-users] Reducing size of a JPEG; can some steps be bypassed? (Magick++)

[Phil Endecott]

Dear Experts,

I have some simple code that uses Magick++ to create 1/2, 1/4, 1/8, 
..-size versions of a JPEG photo.  This is while importing them into a 
photo-browser application and because typically you'll import a large 
number of photos at one time, making it fast is important (quality 
needs to be good but not perfect).

Here are a couple of things that I'm already doing:

1: I use a number of threads working on different images, to exploit 
multi-core processors.

2: I'm loading the original image like this, so that the initial 
decompression doesn't do more work than it has to:

Image i;
i.read(Geometry(...half-size...),filename);

(It took me a long time to work out how to do this, as I was expecting 
to find a constructor that would take the geometry:  Image i(filename, 
new_size);  -  could this be added?)


Here are a couple of further ideas that I've had, but have not yet 
implemented; advice would be appreciated:

3: There is / was a variant of libjpeg that used MMX or SSE 
instructions for improved performance.  But it seems to have been 
ill-maintained and for example it was removed from Debian.  Is anyone 
using something like this with ImageMagick?

4: I'm wondering if it's possible to make things faster by operating at 
a lower level.  I don't think this is possible using Magick++, but 
perhaps it can be done using libjpeg; so it's a bit off-topic here, but 
I'm hoping that there are some JPEG experts reading who can tell me if 
it's feasible.  libjpeg lets you specify scaling of 1/2, 1/4 or 1/8 to 
apply during decompression.  It also lets you retrieve raw data before 
the actual pixel values are computed; either the raw DCT coefficients 
or the data from before the colour-space conversion.  So, can I 
decompress with scaling, get DCT coefficients, and pass them into the 
compressor?  My guess is that I can't because the scaling is done after 
the DCT, but I don't know that for sure.  But perhaps it's possible to 
take a square of 4 DCT blocks and scale them down to a single block?  
Or, perhaps I should do the DCT and just bypass the colour-space 
conversions.  Maybe there is some other library out there that does this.


Any comments about these ideas, or other suggestions for faster 
resizing to 1/2, 1/4... would be much appreciated.

Thanks,  Phil.