Perceptual Image Compression at Flickr

What they did was super impressive, and is a fascinating way to figure out the optimal global compression settings. However, I got my hopes up that perceptual image compression might be about computer vision. At scale, analyzing every image would obviously be prohibitively expensive, but it would be amazing to analyze the detail/edges or color usage (e.g. how much red) in a photo, and tune the q & chroma values accordingly. That, or compress with multiple settings and analyze the resulting images for noise introduced by artifacts.

This is interesting but would be more interesting if they linked to the software that they are using and gave information about the settings used.

Excellent writeup! I'm amazed people signed up to play their test game. Does anyone know if the points you get translate to any sort of other reward, or is a place on the leaderboard more than enough of an incentive to keep playing?

Personally though, I think the linked article about on-the-fly resizing[0] was a more interesting "technical" read.

[0] http://code.flickr.net/2015/06/25/real-time-resizing-of-flic...

I read this looking to see if there was anything we could adapt for mod_pagespeed. Making images smaller without looking worse is, of course, what anyone optimizing images wants. Unfortunately, all they say about how they actually did their perceptual compression is:

    Ideally we’d have an algorithm which automatically tuned
    all JPEG parameters to make a file smaller, but which
    would limit perceptible changes to the image.  Technology
    exists that attempts to do this and can decrease image
    file size by 30-50%.

As far as I know, the only practical way to do this is to do a binary search for the quality level that gives a desired SSIM [1], but this is kind of expensive and I was hoping that they had something cool to tell us here.

[1] If you're interested in getting this kind of improvement and your images are viewed enough that it's worth doing more intensive encoding to get things exactly right, you could do this:

    def evaluator(image, candidate_quality):
      compressed = jpeg_compress(image, candidate_quality)
      return dssim(image, compressed)

    best_quality = binary_search(
      image, desired_dssim, candidate_qualities, evaluator)

Okay, but how is this "perceptual image compression" done?

I wonder if jbig-style pattern matching and substitution (See the Xerox fail) could be used on photos in jpg format.

If the 8x8 tiles in a jpg were lossy deduplicated, the Huffman compression should work better, even if you kept the high frequencies.

http://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres_...

I've been using Flickr for a while now, and am a regular visitor to flickr.com/explore. It has always been so painfully slow to load. Moving to the final photo page is even bigger a task. Sometimes, the thumbnails take two minutes to load, but if I get a direct link to the 4MB image, it downloads in literally 5-10 seconds.

I'm waiting to get home to see if they've done to fix this mess.

Does anyone know what the equivalent imagemagick convert command could be for jpeg and webp? If Firefox supports webp in the near future and IE and Safari support can be done through javascript decompression what could the size reduction with a similar process on webp be compared to jpeg be?

> 15% fewer bytes per image makes for a much more responsive experience.

I wonder how they measured this. Seems like perceived speed gains would be sub-linear with file sizes to me. Particularly as actual transfer speed is often too once you factor in network latency patterns.

This is all impressive and all, but flickr needs to improve their page load times, a lot. I like flickr a lot, and as a paying customer, I'm tempted by 500px every time I visit them.

Took me far too long to realise the screenshot of the comparison test was not interactive. :(

does netflix do this too? or does netflix just get like insane raw source files?