Changing RGB channel intensities using Keras ImageDataGenerator

Question

I have been trying to augment my around 360 images while training a model. This is how the code looks like

img_data_gen_args = dict(rotation_range=90,
                 width_shift_range=0.3,
                 height_shift_range=0.3,
                 shear_range=0.5,
                 zoom_range=0.3,
                 horizontal_flip=True,
                 vertical_flip=True,
                 fill_mode='reflect')
mask_data_gen_args = dict(rotation_range=90,
                 width_shift_range=0.3,
                 height_shift_range=0.3,
                 shear_range=0.5,
                 zoom_range=0.3,
                 horizontal_flip=True,
                 vertical_flip=True,
                 fill_mode='reflect',
                 preprocessing_function = lambda x: np.where(x>0, 1, 0).astype(x.dtype))

image_data_generator = ImageDataGenerator(**img_data_gen_args)
image_generator = image_data_generator.flow_from_directory(train_images_path, 
                                                       seed=seed, 
                                                       batch_size=batch_size,
                                                      class_mode=None,target_size=(IMG_HEIGHT,IMG_WIDTH))  #Very important to set this otherwise it returns multiple numpy arrays 
                                                                        
mask_data_generator = ImageDataGenerator(**mask_data_gen_args)
mask_generator = mask_data_generator.flow_from_directory(train_masks_path, 
                                                     seed=seed, 
                                                     batch_size=batch_size,
                                                     color_mode = 'grayscale',   
                                                     class_mode=None,target_size=(IMG_HEIGHT,IMG_WIDTH))


valid_img_generator = image_data_generator.flow_from_directory(val_images_path, 
                                                           seed=seed, 
                                                           batch_size=batch_size, 
                                                           class_mode=None,target_size=(IMG_HEIGHT,IMG_WIDTH)) 
valid_mask_generator = mask_data_generator.flow_from_directory(val_masks_path, 
                                                           seed=seed, 
                                                           batch_size=batch_size, 
                                                           color_mode = 'grayscale',   #Read masks in grayscale
                                                           class_mode=None,target_size=(IMG_HEIGHT,IMG_WIDTH))  


 train_generator = zip(image_generator, mask_generator)
 val_generator = zip(valid_img_generator, valid_mask_generator)

Apart from these augmentations I would like to use the functionality to change the RGB color intensities to make the model robust in low light conditions as well. Could anybody give a tip how to implement apply_transform function from the ImageDataGenerator class, in addition to what augmentations done in the code above.

Answer 1

You can add a preprocessing_function to the img_data_gen_args dictionary like you do for the mask_data_gen_args dictionary. Inside the function, you can change the RGB color intensities as you like. All you need to do is to make sure that the function takes one argument (Numpy tensor with rank 3), and returns a Numpy tensor with the same shape. Here is the link to the docs. Your preprocessing function would look like so:

def my_preprocessing_function(img):
    # Apply relevant transformation to image here
    return transformed_img