New image and color management library for Python 2+3

Tue Aug 14 14:22:54 EDT 2012

Hello fellow Pythonistas,

I'm looking for co-developers, testers, documentation writers and users
for a new image library I created. The code is available at
https://bitbucket.org/tiran/smc.freeimage

Background story:
I'm working for a company that creates Python based solutions for
libraries -- these large buildings containg millions of books. One major
area of operation is the digitalization of books, prints and manuscripts
from the last millennium. We are a major player in the German speaking
countries with several hundred Terabytes of images on all our installations.

Several years ago I began to write a Python interface for FreeImage
because I was unhappy with the speed and capabilities of PIL. We had to
check and convert several hundred GB of TIFF images to up to 15 sizes
each and store them as JPEG every week. FreeImage offered more features
and speed. Cython made it possible to integrate FreeImage into our
software stack nicely. Later LittleCMS was integrated as color
management solution to convert images from the scanners' color space
into sRGB.

During the development of smc.freeimage I also created a fork of
FreeImage that wraps libjpeg-turbo instead of libjpeg.

In my opinion the project is useful for other people because it has some
benefits over PIL. For example it supports modern file formats that
aren't supported by PIL, for example JPEG 2000, HDR formats and RAW
camera formats. Contrary to PIL it even supports G3 and G4 compressed
TIIFs as well as multipage TIFFs. In some areas it's also much faster
than PIL, especially JPEG performance and NumPy buffer access.

However it's not going to replace PIL as neither FreeImage nor my code
supports drawing, text composition or extended filters.

Excerpt from the README:

Features of FreeImage
=====================

FreeImage wraps mature and widely-used libraries like LibJPEG,
LibOpenJPEG,LibPNG, LibRaw, LibTIFF4, OpenEXR and zlib in a consistent,
well documented and powerful set of APIs.

http://freeimage.sourceforge.net/

Reading of 35 file formats and writing of more than 19 file formats as
of FreeImage 3.15.3, including JPEG 2000, multiple subformats of TIFF
with G3/G4 fax compression and JPEG subsampling.

pixel depths from 1-32 bpp standard images up to formats like RGBAF and
2x64complex.

multi page images

Metadata (e.g. EXIF, IPTC/NAA, GeoTIFF, XMP) and ICC

Color adjustment, conversion and channel processing

Image resizing and rotation

High Dynamic Range (HDR) image processing and tone mapping

RAW camera files

Contrary to PIL it doesn't contain advanced image filters or drawing
functions. FreeImage focuses on file formats

Features of LCMS2
=================

LCMS2 is a color management engine that implements V2 and V4 ICC
profiles up to V4.3. It supports transformation, proofing and
introspection of profiles for a large variety of color formats and targets.

http://www.littlecms.com/

Features of smc.freeimage
=========================

smc.freeimage is developed as part of the closed source Visual Library
framework.

mostly written with Cython with some lines of handwritten C Code and
some Python helpers.

fast, it avoids copying large amounts of data and releases the GIL
whenever possible.

64bit safe, tested on i386/X86 and AMD64/X86_64 systems

thread safe

wraps a large subset of FreeImage features

Compatible with Python 2.6 to 3.3.

Performance
===========

smc.freeimage with libjpeg-turbo read JPEGs about three to six times
faster than PIL and writes JPEGs more than five times faster.

JPEG's restart markers are not compatible with libjpeg-turbo's Huffman
decoder optimization and reduce performance a lot. Please read the
section "Restart Makers" on the page
http://www.libjpeg-turbo.org/About/Performance for more information.

Python 2.7.3
read / write cycles: 300
test image: 1210x1778 24bpp JPEG (pon.jpg)
platform: Ubuntu 12.04 X86_64
hardware: Intel Xeon hexacore W3680 at 3.33GHz with 24 GB RAM

smc.freeimage, FreeImage 3.15.3 standard
 - read JPEG 12.857 sec
 - read JPEG 6.629 sec (resaved)
 - write JPEG 21.817 sec
smc.freeimage, FreeImage 3.15.3 with jpeg turbo
 - read JPEG 9.297 sec
 - read JPEG 3.909 sec (resaved)
 - write JPEG 5.857 sec
 - read LZW TIFF 17.947 sec
 - read biton G4 TIFF 2.068 sec
 - resize 3.850 sec (box)
 - resize 5.022 sec (bilinear)
 - resize 7.942 sec (bspline)
 - resize 7.222 sec (bicubic)
 - resize 7.941 sec (catmull rom spline)
 - resize 10.232 sec (lanczos3)
 - tiff numpy.asarray() with bytescale() 0.006 sec
 - tiff load + numpy.asarray() with bytescale() 18.043 sec
PIL 1.1.7
 - read JPEG 30.389 sec
 - read JPEG 23.118 sec (resaved)
 - write JPEG 34.405 sec
 - read LZW TIFF 21.596 sec
 - read biton G4 TIFF: decoder group4 not available
 - resize 0.032 sec (nearest)
 - resize 1.074 sec (bilinear)
 - resize 2.924 sec (bicubic)
 - resize 8.056 sec (antialias)
 - tiff scipy fromimage() with bytescale() 1.165 sec
 - tiff scipy imread() with bytescale() 22.939 sec

Comparison to PIL (Pros and Cons)
=================================

Pros of smc.freeimage
---------------------

Faster! JPEG performance is about 3 to 6 times faster than PIL, numpy
buffer access is more than 100 times faster and consumes less memory due
to zero copy design.

Modern file formats! smc.freeimage supports JPEG 2000, HDR and EXR high
dynamic range images and raw camera data (RAW).

Full baseline TIFF support! Contrary to PIL smc.freeimage supports all
flavors of baseline TIFF like G3 and G4 compression and multipage TIFFs.

PEP 3118 buffer interface that exports images as 2d or 3d non-contiguous
buffer.

Correct and optimized integration of a color management system
(LittleCMS2)instead of lcms1 integration including caching of optimized
transformations, in-place transformation and introspection of profiles.

Structured metadata access to EXIF, XMP and IPTC information, also
supports fast loading of metadata without loading pixel data.

Lot's of color types! Bitmap (8bit) with 1, 4, 8, 16, 24 and 32 bits per
pixel, (unsigned) int 16 and 32, float, double gray scale, complex, RGBA
16bit and RGBA floats.

Static build support, no need for "make install". You just need a C99
compatible C/C++ compiler, make and nasm (for FreeImage-Turob).

Cons of smc.freeimage
----------------------

Few image filters, no support for complex image filters in FreeImage

Low quality resize filters are slower than PIL's filters

No drawing API for primitives (lines, circles, boxes)

No text drawing support and libfreetype integration.

Still not feature complete and under development.

FreeImage + libjpeg-turbo
=========================

An experimental fork of FreeImage with libjpeg-turbo is available at
https://bitbucket.org/tiran/freeimageturbo

Feel free to contact me if you are interested in the library or want to
get involved!

Christian