Use classified test images from alexnet after training to iterate and improve dataset

I am pre-training a neural network for classifying three classes of images.

My datasets, due to their size, are difficult to filter out and clean up and there are many outliers which are not part of any class.

I would like to filter misclassified images out from the test image set (which is a datastore) based on their score and also set a parameter that classifies only if the scores is high enough (let's say 0.9). If the condition is not met, these images are separated from the dataset into a new folder, for example.

Is there a specific example or a script that would use these [YPred, scores] outputs and correlate them to the images that are falsly predicted in the test image set?

Thanks a lot in advance!

Answer 1

[Saurav Chaudhary]

The one way to deal with this is:

• Removing the outliers from datasets. There are various ways you can do that, refer this documentation linkhttps://www.mathworks.com/help/matlab/ref/rmoutliers.html to find more about it.

• Now using classification algorithm and setting threshold (or score) as 0.9 can seperate out missclassified image and you will have an output vector with images having resemblance > 0.9. You can go through the link given below, to look at an example using [YPred, scores]. https://www.mathworks.com/help/deeplearning/ref/seriesnetwork.classify.html

[Filip Fedorowicz]

Thank you for the reply, the first solution is pretty neat and exactly what I was looking for. I managed to however come up with the same TF vector by a script with for loops. It's good to know that there is an alternative. I basically used to output of [YPred, scores] as a filter. Thanks a lot again.

[Saurav Chaudhary]

Happy to help.