Keras ImageDataGenerator method flow_from_directory

Question

I have a question concerning the flow_from_directory method of ImageDataGenerator. My labels are 2D images so I need to generate augmented data with the same transformation for both images and masks. Reading the doc https://keras.io/preprocessing/image/ looks like flow_from_directory accept only images in a few formats (PNG, JPG or BMP). What about .mhd images or directly numpy array? Can I feed them to this module? Thanks

Answer 1

flow_from_directory is going to read the images using PIL ( see here ), so Keras can read all the formats supported by PIL.

However, I didn't see any method to read both masks and images at the same time, and perform the same transformations. I am afraid that you will have to create your own ImageDataGenerator. It's pretty straightforward if you start with the current code.

Answer 2

Also, you can use ImageDataGenerator for any augmentation tools and then use .flow() instead of .flow_from_directory() . this is a cool way for managing images in numpy array format. chek this out:

train_datagen = ImageDataGenerator(rescale=1/255,
                                   rotation_range=40,
                                   width_shift_range=0.2,
                                   height_shift_range=0.2,
                                   shear_range=0.2,
                                   zoom_range=0.2,
                                   horizontal_flip=True,
                                   fill_mode='nearest')
train_generator=train_datagen.flow(training_images, training_labels)

That's because numpy array isn't a kind of directory. So, convert your data set to numpy array format and then try this out. You will get a true result hopefully.