Was the use of ANN necessary, or simply done because it was cool to do? Because at least in my mind it looks like extracting the text and running a brute force program on it would be faster, no?
If my reading is correct, the neural network was used to OCR the image, extracting the text. From there, it is indeed a simple brute-force string search.
The use of ANN was completely unnecessary, as there are only 3 characters in the crossword puzzle. I wanted to try out the OpenCV ANN implementation though and since it's not a real world application it doesn't really matter.
FWIW that's not a crossword puzzle. As a crossword enthusiast, I would be very interested in the implementation of an AI-backed crossword puzzle solver!
You are correct. However, it is stated that 'KAYAK' appears only once and it seems to me that CTRL+Fing should be the first thing to do when searching for a string in a relatively small text. The next step, had the CTRL+F failed, would have been to search for "KAYAK" vertically/diagonally programmatically.
(on a sidenote, you're counting is incorrect, as it is a palindrome there really is only 4 directions for the word to appear: W-E, N-S, NW-SE, NE-SW)
In Paint.Net I only adjusted the curves (CTRL+SHIFT+M) to get a better contrast. The OCR is done by Tesseract. This solution is all about using the right tool for the job.
Found it in a couple minutes by hand. I started from the position (3,3) and circled the board looking for "Y" surrounded by two A's, and checking for the phrase. Working inward it's only a matter of time before you find it. There are also the borders to check then.