NIST Form-Based Handprint Recognition System (Release 2.2) April 2003 Michael D. Garris mgarris@nist.gov This CD-ROM distribution contains software and data supporting Release 2.2 of the NIST Form-Based Handprint Recognition System. This release is has been packaged for internet distribution and represents what is intended to be the final release of software from this project. The contents of this release should be installed in its entirety. The documentation for Release 2.0 (NISTIR 5959, doc/hsfsys2.pdf) applies to this release with the following exceptions. 1. RELEASE 2.2 INCORPORATES ALL CAPABILITIES AND UPDATES ASSOCIATED WITH RELEASE 2.1 1.1 Release 2.1 was necessitated by the fact that Release 2.0 would not successfully compile and execute using the GNU gcc compiler. More specifically, the implementation of numerical recipes used to perform eigenvector-based feature extraction were not as portable as desired. In this new release, eigenvector computation is supported by a set of portable and more robust routines provided from the CLAPACK mathematical library (www.netlib.org/clapack). As a result, compilation scripts are provided that support GNU gcc and gmake (www.gnu.org) and are targeted, but not limited, to the Linux operating system (www.linux.org). Release 2.1 supported the following UNIX Operating Systems: Linux, SGI IRIX, Sun Solaris, and IBM AIX. 1.2 Release 2.1 was also designed to be compiled to run on computers running the family of Win32 operating systems by first installing the Cygwin library and associated port of GNU tools (www.cygwin.com). The abbreviations used to install these target architectures is listed in the table below. Compilation scripts specifically supporting Open System Foundation (OSF), HP UNIX, and SunOS are no longer provided as these operating systems were not available for testing. However, compilation is now generally possible on those systems that support GNU gcc and GNU gmake by installing the Linux compilation scripts. The porting of the software to other operating systems, compilers, or compilation environments is the responsibility of the recipient. Operating System Installation Assumed Abbreviation Byte Order ------------------------------------------------ Linux lin little endian Win32/Cygwin cyg little endian SGI IRIX sgi big endian Sun Solaris sol big endian IBM AIX aix big endian 1.3 Another change in Release 2.1 had to do with the byte order of binary application data files. Release 2.1 was modified to consistently read and write binary files in big endian byte order. A small number of alternative files are provided in the source code libraries for little endian architectures, in which case data units are swapped so that these computers can operate on, and generate application files in, big endian byte order. The software installation scripts assume the byte orders as listed in the table above and install these alternative files accordingly. Alternative source code files supporting big endian byte order have extension ".big" and those supporting little endian byte order have extension ".lit". If your computer has byte order opposite that listed in the table above, you will have to manually copy the version of these source code files appropriate for your particular computer to files having extension ".c". 1.4 Finally, Release 2.2 incorporates all bug fixes generated for Release 2.1. As this project is no longer active, there will be no further postings of updates. If needed, the entire internet distribution will be updated and the date of last modification posted. 2. RELEASE 2.2 HAS BEEN DOWNSIZED FOR INTERNET DISTRIBUTION As mentioned, this release has been configured for internet distribution. Up to this time, this package was distributed exclusively on CD-ROM and required over 370 Mb. This included all training files containing images of segmented characters along with pattern files containing feature vectors derived from these images. Most of the 370 Mb of storage required on CD-ROM belonged to these files. To make the internet distribution of reasonable size (the gzipped tarball is under 5.5 Mb), all training image files and pattern files have been removed. This has the following implications: * The PNN-based system "hsfsys1" will not execute until its pattern files are computed locally by the recipient. See the "readme.txt" files under "$PROJDIR/weights/pnn/". (Note that the MLP-based system "hsfsys2" will execute using the precomputed weights provided in the internet distribution.) * The MLP classifier used in "hsfsys2" cannot be retrained until its pattern files are computed locally by the recipient. See the "readme.txt" files under "$PROJDIR/weights/mlp/". * Pattern files cannot be computed locally without the files of segmented character images distributed in NIST Special Database 19 (SD19). This database is available for a nominal fee of $150.00 from our distributor: Standard Reference Data srdata@nist.gov Prior to computing pattern files, training files from SD19 must be linked or copied into "$PROJDIR/train/" according to the "readme.txt" file in "$PROJDIR/train/". In conclusion, CD-ROM distribution of this package will no longer continue. The software may now be freely downloaded from the internet via anonymous FTP at sequoyah.nist.gov. Any portion of this package may be used without restrictions. Recipients assume all responsibilities associated with its operation, modification, and maintenance. Much success! Mike Garris mgarris@nist.gov