As for this week, I have been working on the ways that can allow my machine learning model to train with a single picture as well as creating a dataset with facial mask

 

Why with a single picture?

I have recently discovered that one of the issue that caused the prediction to be low is unbalanced classes. The more images that each person have submitted, the ones that has the lesser number of images has a lower prediction rate. It creates a problem of making a validation or test sample as its difficult to have representation across classes in case number of observation for few classes is extremely less.

​

So what can I do?

I have decided to take in a single image to process it 100 times with different image processes. The attributes consist of mirroring image, adding noise, adding face mask, flip, crop, etc. This allow the machine learning model to be highly accurate as it learns the different looks of each person.

​

On top of that, the issue that I have with the image processing which is the face mask augmentation but managed to fix it by manipulating the anchor configs.

​

# ====anchors config

feature_map_sizes = [[33, 33], [17, 17], [9, 9], [5, 5], [3, 3]]

anchor_sizes = [[0.04, 0.056], [0.08, 0.11], [0.16, 0.22], [0.32, 0.45], [0.64, 0.72]]

anchor_ratios = [[1, 0.62, 0.42]] * 5

​

This is an overview of how the flow of the wearing mask process works

​

​

​