Keras ImageDataGenerator 返回具有意外失真的图像

Question

I am experimenting with Keras' ImageDataGenerator() in order to augment my image dataset befor I train a CNN model.

The basic generator object that I've employed is:

datagen = ImageDataGenerator(rotation_range = 30,
                            width_shift_range = 0.2,
                            height_shift_range = 0.2,
                            horizontal_flip = True)

I then produce some augmented data with:

batch_1 = datagen.flow(image_batch, y = labels, 
                       batch_size = len(image_batch),
                       seed = 173)

When I try to look into that batch_1 object, I find pictures like this:

As you can see, there are those parallel colored lines that I cannot explain. No online tutorial on ImageDataGenerator() showed these kind of distortions. Is a CNN trained properly if these images are fed in?

Answer 1

A quick search tells me that it might be because you are setting your rotation_range to 30°, ImageDataGenerator then fills in the empty space between the frame and the image with a continuation of your image's border.

Setting the angle to a multiple of 90° might be a solution, or you can simply have a white border.

Answer 2

I faced the same issue when I used ImageDataGenerator. It is not happening due to rotation_range, It actually happens when we use width_shift_range and height_shift_range. We can use different fill_mode to deal with it. I changed my fill_mode to 'reflect' as my dataset contains leaves.

train_iter = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1./255,validation_split = 0.2,
                                 preprocessing_function = tf.keras.applications.efficientnet.preprocess_input,
                                 rotation_range = 40,
                                 zoom_range = 0.10,
                                 cval = 0.,
                                 width_shift_range=0.2,
                                 height_shift_range=0.2,
                                 shear_range = 0.2,
                                 horizontal_flip = True,
                                 vertical_flip = True,
                                 fill_mode = 'reflect')