I have a bunch of spectograms that I want to feed into a CNN. All the images have a height of 400 and a width of 1000 and look like this:
spectogram
I only want to feed the actual spectogram and not the white parts or the axis description to the CNN so I thought i crop it using the preprocess funtion of the keras ImageDataGenerator.

This is what I came up with:

from tensorflow.keras.preprocessing.image import ImageDataGenerator
import os

path_images = os.path.join('..', 'Datasets', 'spectrums')

def crop_image(image):
    cropped_image = tf.image.crop_to_bounding_box(
        image,
        offset_height = 35, 
        offset_width = 89, 
        target_height = 307, 
        target_width = 717
    )
    return cropped_image

image_generator = ImageDataGenerator(
    rescale=1.0/255,
    preprocessing_function=crop_image
)

train_generator = image_generator.flow_from_dataframe(
    dataframe=data_train,
    directory=path_images,
    x_col="spectrum_filename",
    y_col="simplified_style",
    target_size=(400, 1000),
    batch_size=32,
    class_mode="categorical",
    color_mode="rgb" # add color mode
)

I wanted to train my CNN with this image generator but I got the following error:

ValueError: could not broadcast input array from shape (307,717,3) into shape (400,1000,3)

I thought that I have to adjust target_size to the cropped size, but that just gives me another error:

ValueError: width must be >= target + offset.

All the images have the same size of 400, 1000. I don’t really know what I’m doing wrong. Is the preprocess_function not meant for cropping?

My CNN architecture looks like this (I plan on enhancing it once I have the basic data loading done):

cnn_model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(307, 717, 3)),
    MaxPooling2D((2, 2)),
    
    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D((2, 2)),
    
    Conv2D(128, (3, 3), activation='relu'),
    MaxPooling2D((2, 2)),
    
    Flatten(),
    
    Dense(128, activation='relu'),
    Dropout(0.5),
    
    Dense(6, activation='softmax')  # 6 output neurons for 6 classes
])