This work is licensed under a Creative Commons Attribution 2.5 License.
/*====================================================================*
- Copyright (C) 2001 Leptonica. All rights reserved.
- This software is distributed in the hope that it will be
- useful, but with NO WARRANTY OF ANY KIND.
- No author or distributor accepts responsibility to anyone for the
- consequences of using this software, or for whether it serves any
- particular purpose or works at all, unless he or she says so in
- writing. Everyone is granted permission to copy, modify and
- redistribute this source code, for commercial or non-commercial
- purposes, with the following restrictions: (1) the origin of this
- source code must not be misrepresented; (2) modified versions must
- be plainly marked as such; and (3) this notice may not be removed
- or altered from any source or modified source distribution.
*====================================================================*/
README (10 July 06)
--------------------
gunzip leptonlib-1.37.tar.gz
tar -xvf leptonlib-1.37.tar
1. This tar includes library source, function prototypes,
test program source, and sample images for Linux on x86 (i386)
and AMD 64 (x64), and on OSX (both powerPC and x86).
It should compile properly with any version of gcc
from 2.95.3 onward.
Libraries, executables and prototypes are easily made,
as described in 2, 3 and 4.
When you extract from the archive, all files are put in a
subdirectory 'leptonlib-1.37'. In that directory you will
find a src directory containing the source files for the library,
and a prog directory containing source files for various
testing and example programs.
2. When you compile, object code is by default put into a tree
whose root is also the parent of the src and prog directories.
If you want to change the location of the generated object code,
change the ROOT_DIR variable in the Makefile.
3. To compile and link, just use 'make' in src and prog directories.
It is also possible to make a debug version, as well as one that
builds and uses shared libraries. If you want to use shared
libraries, you need to add the location of the shared
libraries to the LD_LIBRARY_PATH. See the Makefile for details.
4. The prototype files (supplied) can be automatically generated
using xtractprotos. You first have to make this program:
make xtractprotos
Then
make allprotos
Things to note about xtractprotos, assuming that you are developing
in leptonica and need to regenerate the prototype file leptprotos.h:
- xtractprotos is part of leptonica; specifically, it is one of
the programs you can make in the prog directory.
- you can also make it in the source directory, as described above
(see the Makefile).
- xtractprotos uses cpp, and older versions of cpp give useless warnings
about the comment on line 23 of /usr/include/jmorecfg.h. For
that reason, a local version of jmorecfg.h is included that has
the comment elided.
5. Pre-requisites for compilation and linking.
The only dependencies required for using leptonlib are four
libraries that are standard with all linux installations:
libjpeg.a (standard jfif jpeg library, version 6.2 preferred)
libpng.a (standard png library, suggest version 1.2.5)
libz.a (standard gzip library, version 1.1.3 or later)
libtiff.a (standard Leffler tiff library, Rev. 6.0, ca. 1997;
typically version 42 before 1991, or later versions
numbered 3.5.X to 3.7.2)
These libraries (and their shared versions) should be in /usr/lib.
(If they're not, you can change the LDFLAGS variable in the Makefile.)
Additionally, for compilation, the following header files are
assumed to be in /usr/include:
jpeg: jconfig.h
png: png.h, pngconf.h
tiff: tiff.h, tiffio.h
What about jpeglib.h? See item 17 below.
There are also two special interfaces to gnuplot, one that is
programmatic and one that uses a simple file format. To use them,
you need only the gnuplot executable (suggest version 3.7.2);
the gnuplot library is not required.
6. If you want to make programs on other platforms:
(a) Windows
A full download of cygwin provides enough libraries and
programs (e.g., gnu make, gcc, libjpeg, libpng, libz)
to compile the leptonlib libraries in src and make the .exe
execuables in the prog directory. Make the following changes
to the src/Makefile:
(1) remove viewfiles.c from the source list IPLIB_C
(2) choose $CC to remove the -fPIC flag to avoid warnings
'make xtractprotos' in src to generate a windows compatible version
Comment out these targets from the prog/Makefile:
alljpeg2ps
alltiff2ps
maketiles
viewertest
(b) Apple PowerPC
This is big-endian hardware. All regression tests I have
run for I/O and library components have passed on OS-X,
but not every function has been tested, and it is possible
that some may depend on byte ordering. Please let me know
if you find any problems.
Make the following change to the makefiles:
(1) in the src and prog Makefile, change $CPPFLAGS to
define -DL_BIG_ENDIAN
'make xtractprotos' in src to generate a mac-compatible version
7. Unlike many other open source packages, Leptonica uses packed
data for images with all bit/pixel (bpp) depths, allowing us
to process pixels in parallel. For example, rasterops works
on all depths with 32-bit parallel operations throughout.
Leptonica is also explicitly configured to work on both little-endian
and big-endian hardware. RGB image pixels are always stored
in 32-bit words, and a few special functions are provided for
scaling and rotation of RGB images that have been optimized by
making explicit assumptions about the location of the R, G and B
components in the 32-bit pixel. In such cases, the restriction
is documented in the function header. The in-memory data structure
used throughout Leptonica to hold the packed data is a Pix,
which is defined and documented in pix.h.
8. This is a source for a clean, fast implementation of rasterops.
You can find details starting at the Leptonica home page,
and also by looking directly at the source code.
The low-level code is in roplow.c and ropiplow.c, and an
interface is given in rop.c to the simple Pix image data structure.
9. This is a source for efficient implementations of binary morphology.
You can find details starting at the Leptonica home page,
and also by looking directly at the source code.
Binary morphology is implemented two ways:
(a) Successive full image rasterops for arbitrary
structuring elements (Sels)
(b) Destination word accumulation (dwa) for specific
Sels. This has been implemented two ways:
(1) By hand. See the low-level code in fmorphlow.c
From the examples given, you can easily
see how to implement others, and why you'd rather
not do this at all.
(2) Automatically generated. See, for example, the
code in fmorphgen.1.c and fmorphgenlow.1.c.
These files were generated by running the
program prog/fmorphautogen.c.
All results can be checked by comparing with
those from full image rasterops. For the
checking of automatically generated code, see
prog/fmorptest2.c and prog/fhtmtest.c.
Method (b) is considerably faster than (a), which is the
reason we've gone to the effort of supporting the use
of this method for all Sels. We also support two different
boundary conditions for erosion.
Similarly, dwa code for the general hit-miss transform can
be auto-generated from an array of hit-miss Sels.
When prog/fhmtautogen.c is compiled and run, it generates
the dwa C code in fhmtgen.1.c and fhmtgenlow.1.c. These
files can then be compiled into the libraries. To check
the correctness of the automatically generated dwa code
with the rasterop version, see prog/fhmttest.c.
A function with a simple parser is provided to execute a
sequence of morphological operations (plus binary rank reduction
and replicative expansion). See morphseq.c.
The structuring element is represented by a simple Sel data structure
defined in morph.h. We provide several simple ways to generate hit-miss
Sels for pattern finding, in selgen.c.
We also provide a fast implementation of grayscale morphology for
brick structuring elements (i.e., Sels that are decomposable
into linear elements). This uses the van Herk/Gil-Werman algorithm
that performs the calculations in a time that is independent of the
size of the Sels. The low-level code is in graymorphlow.c.
10. This is also a source for simple and relatively efficient
implementations of image scaling, shear and rotation.
Grayscale and color image scaling are done by sampling,
lowpass filtering followed by sampling, area mapping,
and linear interpolation.
Scaling operations with antialiased sampling, area mapping,
and linear interpolation are limited to 2, 4 and 8 bpp gray,
24 bpp full RGB color, and 2, 4 and 8 bpp colormapped
(bpp == bits/pixel). Scaling operations with simple sampling
can be done at 1, 2, 4, 8, 16 and 32 bpp. Linear interpolation
is slower but gives better results, especially for upsampling.
For moderate downsampling, best results are obtained with area
mapping scaling. With very high downsampling, either area mapping
or antialias sampling (lowpass filter followed by sampling) give
good results.
For fast analysis of grayscale and color images, it is useful to
have integer subsampling combined with pixel depth reduction.
RGB color images can thus be converted to low-resolution
grayscale and binary images.
For binary scaling, the dest pixel can be selected from the
closest corresponding source pixel. For the special case of
power-of-2 binary reduction, low-pass rank-order filtering can be
done in advance. Power-of-2 binary expansion is done by
pixel replication.
We also provide 3x and 4x scale-to-gray reduction on binary images
to produce high quality reduced grayscale images.
Conversely, we have special 2x and 4x scale-to-binary expansion
on grayscale images, using linear interpolation on grayscale
raster line buffers followed by either thresholding or dithering.
There are also some depth converters (without scaling), such
as unpacking operations from 1 bpp to grayscale, and thresholding
and dithering operations from grayscale to 1, 2 and 4 bpp.
Image shear has no bpp constraints. We provide horizontal
and vertical shearing about an arbitrary point (really, a line),
both in-place and from source to dest.
There are two different types of general image rotators:
a. Grayscale rotation using area mapping
- pixRotateAM() for 8 bit gray and 24 bit color, about center
- pixRotateAMCorner() for 8 bit gray, about image UL corner
- pixRotateAMColorFast() for fast 24 bit color, about center
b. Rotation of an image of arbitrary bit depth, using
either 2 or 3 shears. These rotations can be done
about an arbitrary point, and they can be either
from source to dest or in-place; e.g.
- pixRotateShear()
- pixRotateShearIP()
The area mapping rotations are slower and more accurate,
because each new pixel is composed using an average of four
neighboring pixels in the original image; this is sometimes
also called "antialiasing". Very fast color area mapping
rotation is provided. The low-level code is in rotateamlow.c.
The shear rotations are much faster, and work on images
of arbitrary pixel depth, but they just move pixels
around without doing any averaging. The pixRotateShearIP()
operates on the image in-place.
We also provide orthogonal rotators (90, 180, 270 degree; left-right
flip and top-bottom flip) for arbitrary image depth. The low-level
code is in rotateorthlow.c.
And we provide implementations of affine, projective and bilinear
transforms, with both sampling (for speed) and interpolation
(for antialiasing).
11. We provide a number of sequential algorithms, including
binary and grayscale seedfill, and the distance function for
a binary image. The most efficient binary seedfill is
pixSeedfill(), which uses Vincent's algorithm to iterate
raster- and antiraster-ordered propagation, and can be used
for either 4- or 8-connected fills. Similar raster/antiraster
sequential algorithms are used to generate a distance map from
a binary image, and for grayscale seedfill. We also use Heckbert's
stack-based filling algorithm for identifying 4- and 8-connected
components in a binary image.
12. A few simple image enhancement routines for grayscale and
color images have been provided. These include intensity mapping
with gamma correction and contrast enhancement, as well as edge
sharpening and smoothing.
13. Some facilities have been provided for image input and output.
This is of course required to build executables that handle images,
and many examples of such programs, most of which are for
testing, can be built in the prog directory. Functions have been
provided to allow reading and writing of files in BMP, JPEG, PNG,
TIFF, and PNM formats. These particular formats were chosen for the
following reasons:
- BMP has (until recently) had no compression. It is a simple
format with colormaps that requires no external libraries.
It is commonly used because it is a Microsoft standard,
but has little else to recommend it. See bmpio.c.
- JFIF JPEG is the standard method for lossy compression
of grayscale and color images. It is supported natively
in all browsers, and uses a good open source compression
library. Decompression is supported by the rasterizers
in PS and PDF, for level 2 and above. It has a progressive
mode that compresses about 10% better than standard, but
is considerably slower to decompress. See jpegio.c.
- PNG is the standard method for lossless compression
of binary, grayscale and color images. It is supported
natively in all browsers, and uses a good open source
compression library (zlib). It is superior in almost every
respect to GIF (which, until recently, contained proprietary
LZW compression). See pngio.c.
- TIFF is a common interchange format, which supports different
depths, colormaps, etc., and also has a relatively good and
widely used binary compression format (CCITT Group 4).
Decompression of G4 is supported by rasterizers in PS and PDF,
level 2 and above. G4 compresses better than PNG for most
text and line art images, but it does quite poorly for halftones.
It has good and stable support by Leffler's open source library,
which is clean and small. Tiff also supports multipage
images through a directory structure. See tiffio.c
- PNM is a very simple, old format that still has surprisingly
wide use in the image processing community. It does not
support compression or colormaps, but it does support binary,
grayscale and rgb images. Like BMP, the implementation
is simple and requires no external libraries. See pnmio.c.
Here's a summary:
- All formats except JPEG support 1 bpp binary.
- All formats support 8 bpp grayscale and 24 bpp rgb color.
- All but PNM support 8 bpp palette.
- PNG and PNM support 2 and 4 bpp images.
- PNG supports 2 and 4 bpp palette, and 16 bpp without palette.
- PNG, JPEG and TIFF support image compression; PNM and BMP do not.
Use prog/ioformats for a regression test.
We also provide wrappers for PS output, from the following
sources: binary Pix, 8 bpp gray Pix, 24 bpp full color Pix,
JFIF JPEG file and TIFF G4 file, all with a variety of options
for scaling and placing the image, and for printing it at
different resolutions. See psio.c for examples of how to output
PS for different applications. As an example of usage, see
prog/converttops.c for a general image --> PS conversion for printing.
We provide colormap removal for conversion to 8 bpp gray or
for conversion to 24 bpp full color, as well as conversion
from RGB to 8 bpp grayscale. We also provide the inverse
function to colormap removal; namely, color quantization
from 24 bpp full color to 8 bpp palette with some number
of palette colors. Several versions are provided, that all
use a fast octree vector quantizer. For a high-level interfaces,
see pixConvertRGBToColormap(), pixOctreeColorQuant() and
pixOctreeQuant().
A function can be called to display an image programmatically on
an X display using xv. If necessary to fit on the screen,
the image is reduced using scale-to-gray for readability.
14. Simple data structures are provided for safe and
efficient handling of arrays of numbers, strings, pointers,
and bytes. The pointer array is implemented in three ways:
as a stack, a queue, and a heap (used to implement a priority
queue). The byte array is implemented as a queue. The string
arrays are particularly useful for both parsing and composing text.
Generic lists with doubly-linked cons cells are also provided.
15. Examples of programs that are easily built using the library:
- for plotting x-y data, both programmatic and file
interfaces to the gnuplot program (e.g., plotprog
and plotfile) and with output to X11, png, ps or eps.
- a simple jbig2-type classifier, using various distance
metrics between image components (correlation, rank
hausdorff); see prog/jbcorrelation.c, prog/jbrankhaus.c.
- a simple color segmenter, giving a smoothed image
with a small number of the most significant colors.
- a program for converting all tiff images in a directory
to a PostScript file, and a program for printing an image
in any (supported) format to a PostScript printer.
- converters between binary images and SVG format.
- a bitmap font facility that allows painting text onto
images. We currently support one font in several sizes.
The font images and postscript programs for generating
them are stored in prog/fonts/.
- a binary maze game lets you generate mazes and find shortest
paths between two arbitrary points, if such a path exists.
You can also compute the "shortest" (i.e., least cost) path
between points on a grayscale image.
16. A serious deficiency of C is that no standard has been universally
adopted for typedefs of the built-in types. As a result,
typedef conflicts are common, and cause no end of havoc when
you try to link different libraries. If you're lucky, you
can find an order in which the libraries can be linked
to avoid these conflicts, but the state of affairs is aggravating.
The most common typedefs use lower case variables: uint8, int8, ...
To avoid conflicts, we previously typedef'd the builtins with
UINT8, INT8, .... However, our convention conflicted with the
one in the jpeg library, specifically the file jmorecfg.h:
e.g., we used
typedef int INT32;
whereas jmorecfg.h uses
typedef long INT32;
Consequently, it was necessary to distribute a version of jmorecfg.h
with the typedefs elided (!)
The png library avoids typedef conflicts by altruistically
appending "png_" to the type names. Following that approach,
Leptonica now appends "l_" to the type name. This should avoid
just about all conflicts. In the highly unlikely event that it doesn't,
here's a simple way to change the type declarations throughout
the Leptonica code:
(1) customize a file "converttypes.sed" with the following lines:
/l_uint8/s//YOUR_UINT8_NAME/g
/l_int8/s//YOUR_INT8_NAME/g
/l_uint16/s//YOUR_UINT16_NAME/g
/l_int16/s//YOUR_INT16_NAME/g
/l_uint32/s//YOUR_UINT32_NAME/g
/l_int32/s//YOUR_INT32_NAME/g
/l_float32/s//YOUR_FLOAT32_NAME/g
/l_float64/s//YOUR_FLOAT64_NAME/g
(2) in the src and prog directories:
- if you have a version of sed that does in-place conversion:
sed -i -f converttypes.sed *
- else, do something like (in csh)
foreach file (*)
sed -f converttypes.sed $file > tempdir/$file
end
If you are using Leptonica with a large code base that typedefs the
built-in types differently from Leptonica, just edit the typedefs
in environ.h. This should have no side-effects with other libraries,
and no issues should arise with the inclusion order of the Leptonica
libraries for the linker.
For compatibility with 64 bit hardware and compilers, where
necessary we use the typedefs in stdint.h to specify the pointer
size (either 4 or 8 byte). This may not work properly if you use a
compiler before gcc 2.95.3.
17. For C++ compatibility, we have included a local version of
jpeglib.h (version 6b), with the 'extern "C"' macro added.
The -I./ flag includes this local file, rather than the one
in /usr/include, because of the order of include directories
on the compile line. jpeglib.h will by default include the
locally-supplied version of jmorecfg.h.
18. Leptonica provides some compile-time control over messages
and debug output. Messages are of three types: error,
warning and informational. They are all macros, and
are suppressed when NO_CONSOLE_IO is defined on the compile line.
Likewise, all debug output is conditionally compiled, within
a #ifndef NO_CONSOLE_IO clause, so these sections are
omitted when NO_CONSOLE_IO is defined. For production code
where no output is to go to stderr, compile with -DNO_CONSOLE_IO.
19. If you use Leptonica with other systems, you have three
choices with respect to the Pix data structure. It is
easiest if you can use the Pix directly. Next easiest is to
make a Pix from a local image data structure or the unbundled
image parameters; see the file pix.h for the constraints on the
image data that will permit you to avoid data replication.
By far, the most work is to provide new high-level shims from some
other image data structure to the low-level functions in the library,
which take only built-in C data types. It would be an inordinately
large task to do this for the entire library.
20. New versions of the Leptonlib library are released approximately
monthly, and version numbers are provided for each release.
A brief version chronology is maintained in version-notes.txt.
The library compiles without warnings with either g++ or gcc,
but you will get warnings with the -Wall flag.