I am using keras for an image recognition Sequential CNN model. I have about 300,000 images 424×424, but only the inner 168×168 is relevant.

I am currently using the following code to load the images before training the model, but this just resizes to 168×168 instead of cropping:

from keras.utils import image_dataset_from_directory
train_ds = image_dataset_from_directory(
    train_dir,
    validation_split=0.2,
    subset='training',
    color_mode='grayscale',
    seed=12345,
    image_size = (168,168),
    batch_size=64
)

Is there an easy way to crop out the central 168×168 from train_ds?
I know I can crop the original files easily, but I was hoping there was a way to crop within the code, so I can play around with different sizes easily.