Dice Image Number detector. (Machine Learning)

What we need is a "Dice" detector. You need to create an app that accepts url get requests with an image as a parameter. You should be able to call the app like this: getDiceNumbers?image=[login to view URL] The image will always be , 12 , 6 sided dices with dark background. For example: Here is images data: [login to view URL] (This contains 200 test images). We will need to generate an image similiar to [login to view URL], after each request and serve it via Flask or something, so we can access it like this: showImage?name=[login to view URL] The program should be able to detect, each individual dice numbers, and return the get request with a json object like this, OR give errors on dices. { dice_amt_found: 12, (Amount of dice the program has found) generatedImage: [login to view URL] dices { dice_1: 1, dice_2: 3, dice_3: 4, dice_4: error, //dice is flipped or something .... } } Must: - We must use ML, because openCV will not be 100% detection rate, and its not always correct. - The app MUST have a detection rate of 100%. - We must be able to differentiate on error dices. (The ones that are for example flipped to the side, like image #1 from dataset). - The app should detect the dices individual pips from the dice, even though the dices touch. - The app should give error if there was error on 1 of the dices. (Shows 2 sides like on here [login to view URL]) - provided request Image size should not matter (Should be able to accept high res full HD image and 720p as well.)