If image has (28,28,3) shape, how do i convert it to (28.28,1)?

Question

I'm using the mnist dataset of handwritten digits and I am trying to predict a digit that I wrote. The problem is that my digit is of shape (28,28,3) and the expected shape for my neural network is (28,28,1). How can I convert it?

my code:

import tensorflow as to
from tensorflow import keras
from keras.datasets import mnist
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import cv2

data = mnist.load_data()
(x_train, y_train), (x_test, y_test) = data

classes = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']

x_train = x_train / 255
x_test = x_test / 255

model = keras.models.Sequential()
model.add(keras.layers.Flatten(input_shape=(28,28)))
model.add(keras.layers.Dense(128, activation='relu'))
model.add(keras.layers.Dense(10, activation='softmax'))

model.compile(
     optimizer='adam',
     loss='sparse_categorical_crossentropy',
     metrics=['accuracy']
)

model.fit(x_train, y_train, epochs=7)
test_loss, test_acc = model.evaluate(x_test, y_test)
print('\nTest Loss:', test_loss)
print('Test accuracy:', test_acc)

img = Image.open("7.jpg").convert('L')
img_array = cv2.imread('7.jpg')
new_array = cv2.cvtColor(img_array, cv2.COLOR_BGR2GRAY)
new_array = cv2.resize(new_array, (28,28))
print(new_array.shape)
print(x_test[0].shape)
plt.imshow(new_array, cmap='gray')
plt.show()

predictions = model.predict(new_array)

plt.grid(False)
plt.imshow(new_array, cmap='gray')
plt.title("Prediction: " + classes[np.argmax(predictions)])
plt.show()

Adamsite

You want to reduce your color space to black-and-white? – Tollman

If you want to reduce the color depth from 24-bit (3 channels) to 8-bit (1 channel) convert the image to grayscale using cvtColor. If you want to further reduce the info to black and white (1 channel) threshold the gray output. – Packthread

After using cvtColor, it says: expected flatten_input to have 3 dimensions, but got array with shape (28, 28) – Adamsite

It's hard to guess where the error is without ANY of your code. My guess is that you are giving a bad mat to cvtColor and/or giving the wrong conversion code. – Packthread

Just added my code. thanks for helping 👍🏻 – Adamsite

Answer 1

Assuming img has shape of (28, 28, 3) you can do this:

gray = cv2. cvtColor(img, cv2.COLOR_BGR2GRAY)
gray = np.expand_dims(gray, 2)

This will convert it to a shape of (28, 28, 1)

Stockdale

Answer 2

Even though you want a (28,28) shape, tensors need to have 3 dimensions, so you have to expand that to (28,28,1). Something like this should be enough:

new_array = cv2.resize(new_array, (28,28,1))

Or expand it using new_array=new_array[..., None].

Henka

after using: new_array=new_array[..., None] I get an error: expected flatten_input to have shape (28, 28) but got array with shape (28, 1) – Adamsite

What about the first option I suggested? Also where's the error coming from? – Henka