I am working on a model to detect images. There are 2 classes, real and generative (fake). I can't get higher than 0.85 f1_score. Any recommendations how to improve the score?
The data set contains 4000 real images (4000, 1200) and 2000 fake images (2000, 1200)
So far I tried several ensemble methods like XGBoost, KNeighbours, SVC and Random Forest, and deep learning. Unfortunately, I can't use pretrained models like VGG or ResNet as the images size must be for most of them higher than 32 and by converting to (40, 30, 1) or (20, 20, 3), I get an error.
As the data set is imbalanced, I tried SMOTE oversampling and data augmentation and it improves a bit. Also, I tried Robust, MinMax and Standard Scaler. But so far, I can't get higher.
I would appreciate any recommendation to improve:
- Any other preprocessing technique?
- Any other model? I read GANs could work well but I never used them before, and I am trying to fing a starter notebook.
