Published
Reading Time3 min read
We recently worked with the team at Cloudinary to help build and evaluate a
better image compression and quality measurements. The results of Cloudinary’s
work on Scale were very insightful and we wanted to share it broadly to
demonstrate how more companies can leverage human judgments to build high
quality features.
Without further ado,
Tal Lev Ami, CTO
of Cloudinary, wanted to write about how they used Scale:
Here at Cloudinary, we provide a cloud-based tool that enables our users to
compress images and video for their websites and apps. Our goal is to
preserve the visual integrity of the content, but deliver the smallest file
size to any device or browser to ultimately optimize website performance and
end user satisfaction.
One of the hallmarks of the Cloudinary solution is the ability to automate
many functions of image compression, so that developers don’t have to spend
time tweaking each photo and making multiple copies of different sizes and
resolutions to fit every possible scenario. Compression algorithms can be
tricky because they’re trying to make changes that have the smallest visual
impact, but different images can react differently to compression.
As we were developing the algorithm for our “q_auto” capabilities — which
strikes a balance between visual quality and file size — we needed to test
it to understand how the resulting images compared to the human eye. Enter
Scale.
Many image compression formats — like JPEG 2000 and JPEG XR — have been
tweaked to score well on particular metrics, such as peak signal-to-noise
ratio (PSNR). But these don’t always correlate with human perception on
image quality.
We leveraged Scale to compare pairs of images and give us perspective on
which image was liked most by humans. With Scale, we did a variety of tests,
comparing several formats, including WebP, JPEG 2000, JPEG XR (lossy) Lepton
(MozJPEG, recompressed with Lepton), FLIF, BPG, Daala, and PNG8
(pngquant+optipng). We also were able to get input on the difference between
the uncompressed original image vs. a compressed version.
Scale enabled us to create A/B comparisons that were viewed by human
observers. We submitted over 4,000 image comparisons to Scale, sending at
least four independent Scale requests for each pair of image. This resulted
in at least eight actual human comparisons for each pair of images. The
outcome of these comparisons were evaluated beside other perceptual metrics
such as PSNR, Google’s Butteraugli, DSSIM (Structural (Dis)Similarity) and a
new metric Cloudinary developed called
SSIMULACRA (Structural SIMilarity Unveiling Local And Compression Related
Artifacts).
The results showed that overall, PSNR is “correct” in only 67 percent of the
cases. Butteraugli gets it right in 80 percent of the cases, and DSSIM in 82
percent of the cases. Our new metric, SSIMULACRA, agrees with human
judgments in 87 percent of the cases. Looking just at the high-confidence
human judgments, we found about 78 percent agreement for PSNR, 91 percent
for both Butteraugli and DSSIM, and almost 98 percent agreement for
SSIMULACRA. You can
read more about SSIMULACRA and these results on the Cloudinary blog. Or if you want to give it a try:
SSIMULACRA is free and open-source software!
The results of Scale comparisons gave us useful data points to validate our
metrics and provided more insights into the compression benchmarks we are
running and the comparison of various image formats. And from these insights
we were able to improve our visual perception metrics and fine-tune our
“q_auto” functionality so we know how aggressively we can compress images.
Through this process we were impressed not only by the useful data points
derived from the Scale, but also the great support we got from the company
and the product’s ease-of-use, all which came at a reasonable price.