Attempting to create a code which will run in situ with a ghost imaging experiment to identify an object as soon as possible, without it fully being formed.
There are two folders with images:
(1) The "Training Data.zip" folder currently holds 2000 images, 1000 variations of three and 1000 variations of four. The two folders are called in the "number_recognition_data_input.py" code.
(2) The "Predict against.zip" folder contains falsely created images of 3's and 4's that have varying levels of noise on the image. This is used in the "number_recognition_num_predict.py" code.
And currently there are three program files included:
(1)The first file is the "number_recognition_data_input.py" which turns the training data (3's and 4's) into arrays and pickles it into an X and y pickle. I have 15000 of each 3 and 4 where the actual number is shifted, rotated and dialated so there should be no issue with the data being the same that I see.
(2)The second file, "number_recognition_cnn.py" is the cnn that seems to overfit the data. This is where the major issue lies I believe, No combination of cnn seems to not produce an accuracy of over 90% within three epochs.
(3)The "number_recognition_num_predict.py" file is to purely test against falsely created data to see if the cnn can accurately identify a 3 from a 4.
This is the journal article I have found that might be able to help: https://www.nature.com/articles/s41598-017-18171-7