Depending on how your target application displays its windows, you can possibly dig down through it to get at the real text and caption of the window instead of needing to capture screens.

Take a look at the included minispy.pl program and especially the FindWindow() routine - you will need a bit of knowledge about GDI or generally windowing APIs though, to make sense of the data returned by it.

Otherwise, if you have to stay on the graphic route, consider using PDL - I think PodMaster has a PPM package for Win32.

I'm at work now, where I use Win32::GUITest to automate a stubborn application, and here is a modified version of the spy-- program that I used to look at the current selection in a listbox. The stubborn program displays its status by filling a listbox with all possible status values and then selecting the active status...

# $Id: guitest-spy.html,v 1.1 2001/09/19 23:31:36 erngui Exp $
#
# MS has a very nice tool (Spy++).
# This is Spy--
#
use Win32::GuiTest qw/FindWindowLike GetWindowText GetClassName IsWind
+owEnabled GetWindowID
    GetChildDepth GetDesktopWindow GetComboContents GetListContents Ge
+tListText SendKeys /;
use Win32::API;

for (FindWindowLike(0,qr/^Explorer -/)) {
    SendKeys("%ao");
    print "Looking for dialog\n";
    my @dialog = FindWindowLike($_,qr/Optionen/,"");
    print "Looking for listbox\n";
    my @listbox = FindWindowLike($dialog[0],"","",752);

    while (1) {
      print "Getting current selection\n";
      my $i = SendMessage($listbox[0], LB_GETCURRSEL,0,0 );
      print $i;
      print "Getting current selection text\n";
      my $t = GetListText($listbox[0],$i);
      print $t,"\n";
      sleep 1;
    };
}
[download]

Corion, ++ and thanks for the example! We're hoping to start writing some GUI test automation here, could I steal^H^H^H^H^H copy that code for our internal use? It would sure be nice to be able to develop in-house what CA wanted $20K+ for, and that code will save me at least a few hours.

--
Spring: Forces, Coiled Again!

Thanks, but Application is running within a windows frame utilizing some Java SunAwtCanvas... anything within it is not visible to me from Perl. only graphically. Found the Win32::ActAcc module, here is the Perl code I used to detect any "lasting change" over a 2 second period... It appears very handy for working with other applications...

use Strict;
use Win32::OLE;
use Win32::ActAcc;
use Win32::GuiTest qw(:ALL :SW);
 
 my @win;
 my %Wnm;
 my %VW;

 Win32::OLE->Initialize();
 while ( 1 ) {
  @win = ();
  @win = FindWindowLike(); 

  #*# report any "lost" windows:
  for $win ( @win ) { delete $Wnm{$win}; }

  if ( scalar keys %Wnm) {
    print "Lost:\n";
    my @Lost = sort keys %Wnm;
    for (@Lost) { print "$Wnm{$_}\n"; delete $Wnm{$_}; }
    print "\n";
  }

  for $win ( @win ) { dumpwin($win); }

  print "\n----\n";
  sleep(2);
 }
exit;

sub dumpwin {
  my $win = shift;
  print "Null handle\n", return unless ($win); 

  my $w_text   = $win . ' ' . GetWindowText($win); 
     $w_text  .= " class="  . GetClassName($win);
  my $w_state  = ' V=' . IsWindowVisible($win) . ' E='; 
     $w_state .= IsWindowEnabled($win); 
  $Wnm{$win}  = $w_text . ' ' . $w_state;

  my $ao = Win32::ActAcc::AccessibleObjectFromWindow($win);
  my $aatxt = '';
  $aatxt = $ao->describe() if defined $ao;
  $w_state .= ' ' . $aatxt;

  if ( exists $VW{$win} ) {  
    return if ( $VW{$win} eq $w_state );
  }
  $VW{$win} = $w_state;

  print $win,      GetWindowText($win);
  print " class=", GetClassName($win);
  print " vis=",   IsWindowVisible($win); 
  print " en=",    IsWindowEnabled($win); 
  print "\n ",$aatxt, "\n"; 
}
[download]

This problem is called "image registration" or "image stitching" or "image mosaicing". There are plently of papers written on the subject, most of which don't use AI techniques. The most widely used method is getting the Fourier transform of both images and comparing them; but this only works if there is no change in rotation between the images. I, personally, have written a paper to do just that using Zernike moments. Many other approaches exist, each with it's ups and downs.

One ugly and fun way would be to use the regex engine. 1024*786 is less than 1 MB so it could be fast even. (Assuming 8 bpp)

Just insert .{ screen dump width - expected image width } after every expected image width characters (but not at the end).

For real image searching, there are a lot of interesting tests. Juerd mentioned using the regex engine, which has its merits. There are other sparse comparisons which help find exact pixel matches of various swatches. Another poster mentioned image mosaicing methods like least squared errors, but those are better for situations where the alignment, rotation, scale, coloring of two images are not identical.

Otherwise, please please please dig through the window manager to verify text strings, not pixels. Any change in the test user's Appearance preferences, or operating system version, or even a different video driver, would totally blow all your tests.