Data Compression

Hi Robert

for me and any others who are more interested in the answer to a problem than capable of following the maths in the middle, are you able to improve on my empirical blundering (obviously taking an image where there is some detail worth keeping) -

Taking an image that has to be scaled down to about 33% for forum display, I find Lanczos 3 compression has a marginal edge on the image reduced to 768H over Bicubic or Mitchell. Comparing original at 67% & reductions at 200%, Lanczos still has an edge. Comparing the original at its full size and the reductions at 300%, there is virtually nothing to choose between the compressed images.

But there is still a little to be gained on going back over the Lanczos reduced image using 'Local contrast enhancement'. LCE, being done with controlled values, seems much more successful than 'bicubic with sharpen'.

I suspect that there is then no universal answer to the last bit, ie whether in general it is best to use the more subtle sharpening techniques before or after compression? (getting a better camera, better lens and always using a tripod not being a realistic option)

[robski]

Chris

Re-sampling is more than slightly off topic for this thread.

The main problem with up or down sampling is which bits do you throw away or expand. The problem is made worse if the re-sizing is not a multiple of the original. i.e your scaling by 1.4567 %. Beat patterns can occur in the image due to pixels within image information changing in position or being added or dropped. After rescaling some techniques apply a dithering to hide regular beat patterns.

I am afraid I don't have a great deal of experience of re-sampling images. A lot of my postings are 100% crops or down-sampled using bi-linear.

Maybe Nigel Blake can offer some input on this topic.

[robski]

Just in case a few of you were wondering what beat patterns are about I thought I would illustrate the effect.

I made an image with alternate black and white 10 pixel wide bars. I then scaled the image by 5.2% using nearest neighbour method ( upper half of attachment) and a method that looks at it's surrounding pixels and dithers (lower half of attachment). If you look in the upper half you will notice that every so often one of the white or black bars is slightly wider. From a distance the eye will pick this up as an interference pattern. In the lower half the dithering has reduced this effect.

[robski]

I have taken the trouble to tidy most of this information up and posted here

Any feedback, comments or corrections gratefully received.