steps_taken.txt

Hi!  As mentioned, the purpose of this exercise is to teach users to do basic computer vision in Python.

First things first, I needed our program to recognize text from an image.  We needed to recognize score for our NN to work.

I started with Selenium, ran the program, retrieved the score using PIL and OpenCV, and utilized pytesseract.
It was atrocious.  Firstly, 8's were shutdown as 0s.  6's could be G's.  5's were S's.  2's were z's!
I did all sorts of operations.  Preprocessing, image enlargening, image-smallering?
Making the image smaller actually helped pytesseract, but I still had too much error for comfort.

As such, I wanted to do custom character recognition for greater accuracy.  Problem was,
google's dino game uses a tricky and rather specific font that I couldn't identify, even after extensive 
searching.

So I embarked on a font hunt, and ended up finding their sprite sheet here:
https://github.com/wayou/t-rex-runner/blob/gh-pages/assets/offline-sprite-2x.png

Not exactly what I wanted, but it was good enough to work with!
From here, I needed to extract only the digits.

The first couple of lines in the program do just that.
After which, all I did was some basic pre-processing and contour application to grab the templates for our digits.

Hah.  Okay I say basic but it would have been tons harder without Adrian Rosebrock's beautiful tutorials.  
If you'd like to try your hand at auto number/text detection, check out their link below:
https://www.pyimagesearch.com/2017/07/17/credit-card-ocr-with-opencv-and-python/


Basically, contour-ing grabs the outlines of the digits, then we apply boxing and stuff to get the full digit.

After grabbing our digit templates, we grab the region of the screen where the score appears,
pre-process it, apply contours again, and check which template best matches an individual number in our score.
(And yes, we could have done all this to the high-score image instead.  It would not be as cool though..)

But Kablam, a FAR superior accuracy to PyTesseract!



Okay, so we've gotten a reliable means of tracking the score.
That leaves actual object detection.  Cacti don't dodge themselves.

Using our screenshot, we grab a still-background image for our regular frame.  Preferrably the region
right of our 'saur.  Be sure to exclude the ground from your image.  Let the pre-processing commence.

Next step, we compare a current frame with our first, then apply contours and filling to track the moving
object.  You can actually do real-life motion tracking later with a camera if you prefer!  Definitely check out:
https://www.pyimagesearch.com/2015/05/25/basic-motion-detection-and-tracking-with-python-and-opencv/



Recap:
By the end of this exercise, we learned how to apply contours to recognize digits, match objects, and
detect basic motion.  We have experience in detecting objects on screens, applying pre-processing, and 
hopefully, we gained a thing or two with tools like OpenCV, PIL, imutils, and Selenium.