You don't remove images that are overfit, you adjust parameters like model size, add error functions for exact reproductions training data, etc.

You can deduplicate images that are overfit if you have hundreds of them in your data set, which makes it much less likely overfitting happens.