CODEWITHSUNDEEP
numpyopencvimage-processingkerascomputer-vision
Suhail Gupta
Suhail Gupta

Reputation: 23256

Bad Argument error as I try to add noise to an image with Keras preprocess function

To perform image-preprocessing, I was trying to use the ImageDataGenerator class from Keras. Here is how used it:

    data_generator = keras.preprocessing.image.ImageDataGenerator(
        rotation_range = 60,
        width_shift_range = 0.1,
        height_shift_range = 0.1,
        brightness_range = [0.5, 1.5],
        shear_range = 0.01,
        zoom_range = [0, 1],
        horizontal_flip = True,
        vertical_flip = True,
        preprocessing_function = preprocess_other
   )

The preprocessing_function attribute has been assigned a function named preprocess_other as defined below:

    def preprocess_other(image):
        flip = np.random.random()
        if flip > 0.5:
            # Add noise
            blank_image = np.zeros(image.shape, np.uint8)
            cv2.randn(blank_image, 0, 5)
            noisy_image = cv2.add(image, blank_image)
            return noisy_image
        else:
            # Return the original image
            return image

The role of this function is to add noise to an image with a probability of 0.5.

As I start the training process (training the CNN), it works for a few seconds but fails due to some error with the preprocess_other function with an error saying:

error: OpenCV(3.4.3) /io/opencv/modules/core/src/arithm.cpp:683:
error: (-5:Bad argument) When the input arrays in 
add/subtract/multiply/divide functions have different types,
the output array type must be explicitly specified in function 
'arithm_op'

I debugged but could not understand the reason for it. Am I trying to add the noise in an incorrect way? How could I correct this error?

Upvotes: 1

Views: 106

Answers (1)

Berriel
Berriel

Reputation: 13611

The problem is that image and blank_image have different types.

You can change:

blank_image = np.zeros(image.shape, np.uint8)

to:

blank_image = np.zeros(image.shape, image.dtype)

or to:

blank_image = np.zeros_like(image)

Upvotes: 1

Related Questions