I suppose you might break it down into separate steps. It's not elegant, but it does seem to work:

my $inputText = "word1, word2, #18abcdefgh ,word4";

print "before: $inputText\n";

if ( $inputText =~ /#(\d)/g )
    {
    my $n1 = $1;
    if ( $inputText =~ /(\d{$1})/g )
        {
        my $n2 = $1;
        if ( $inputText =~ /(.{$1})/g )
            {
            print "(1) = $n1\n";
            print "(2) = $n2\n";
            print "(3) = $1\n";
            }
        }
    }

print "after: $inputText\n";
[download]

Update: on re-reading the question, the code above only extracts the bits you are after; it does not do the string substitution. I suppose there a number of ways you might do that; one way that springs to mind is to build a new string rather trying to substitute (you may need m//gc for that to stop the match operator resetting the position within the string when a match fails). Something like this:

use strict;

my $inputText = "word1, word2, #18abcdefgh ,word4 ";
$inputText .= "word5, word6, #212abcdefghijkl ,word7\n";

print "before: $inputText\n";

my $newstr;
{
    if ( $inputText =~ /\G#(\d)/gc )
        {
        my $n1 = $1;
        if ( $inputText =~ /\G(\d{$1})/gc )
            {
            my $n2 = $1;
            if ( $inputText =~ /\G(.{$1})/gc )
                {
                print "(1) = $n1\n";
                print "(2) = $n2\n";
                print "(3) = $1\n";
                $newstr .= "\<Binary block: $n2 bytes\>";
                }
            }
        }
    elsif ( $inputText =~ /\G([^#]+)/gc )
        {
        $newstr .= $1;
        }
    else
        {
        last;
        }
    redo;
}

print "after: $newstr\n";
[download]