Image noise matches image text

Question

So I've been reading up on machine learning using TensorFlow and Keras, I've been trying to setup a dataset using some custom images and trying to learn the script to recognize the text while filtering out the noise, but the issue is that the noise color is the same and the text color which results in filtering out everything.

I'm not asking to get spoonfed, I just simply want pointers to the best way to solve/train the script to solve the text on the images.

What I'm looking for is to get the script to read on screen and calculate the word hidden in the image and print the result in the command line.

There is no sample code since everything before was a failure and not actually what I was looking for.

Album link for Imgur

Answer 1

before you start your project, you should check the "Quality" of your data, and it's value for the real-time application or your project, and if the images is not readable by humans easily, then it's a little bit wrong to train with this kind of data because humans tend to be very good at recognizing things visually, and a lot of times the "Bias error" is taken depending on that.
Any way if you are AIMING to a read a noisy text from images, you can try this tips which are taken from an online course on coursera named: "Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization" By prof.Andrew Ng:

1 - Train your data on normal and clear text-image images, because that will let your algorithm learn a lot of features and Initial proprities from the clear images, like the shape of the letters for example, and you may be surprised by the results sometimes.

2 - Let your DEV set (and your test set) contains a lot of noisy text-image images, so that you can check how your algorithm is really doing on the data you really care about(your AIM)

3 - Changing the dataset may be a little bit Difficult, but if the noise makes a part of the foto in a way that even for a human it's difficult to read, this dataset may be useless.(Not always) so you can bring a noise-clear images and mix them with a noise images(make a new dataset from the old one) in a way that it remain realistic and not much robotic and then train the algorithm on your new data.

Building a good ML algorithm or app depends in the first place on your data, and those are just some notations that may help you thinking about the problem in another way.