What Mordecai said.
However, also, possibly, if the text is always the same size, same font, same color against the same color background, and you really fancied some work, you could find the top left point of the string of letters/words (first letter of the first word), by using some of the GD (from Boutell) functions to read the color at certain coordinates of the image, until you found the first letter, then run some routine that identifies letters from looking around within the same vicinity for the right color in certain offset coordinates (whatever are the letter and not background) and then kind of move right finding each letter.
OCR/character/word recognition packages may also have a thing where looking at what has been found already, they can guess by frequency/probability what letter/word is next according to their records. I think. Something like that.
