If you’re using opencv and/or tifffile to prepare your dataset for training deep learning models, beware the difference between the imread function in both libraries, i.e., cv2.imread and tifffile.imread. It makes a huge difference, especially if you have custom, multi-channel images.

Let’s experiment with an example custom tif image with dimensions 224x224x4 below.

# import necessary modules

import cv2
import tifffile
import matplotlib.pyplot as plt
%matplotlib inline


# read image with tifffile

image_tifffile = tifffile.imread('example.tif')
print(image_tifffile.shape) # prints 224x224x4


(224, 224, 4)

# read image with cv2
image_cv2 = cv2.imread('example.tif', cv2.IMREAD_UNCHANGED)
print(image_cv2.shape) # also prints 224x224x4

(224, 224, 4)

Looks great, the same image read with both libraries are read and have the same size. However, there is a twist. Let’s use matplotlib and plot all four slices for the image read by both libraries.

# plot all four slices for both images read above
fig, ax = plt.subplots(2, 4)
for i in range(4):
    ax[0, i].imshow(image_tifffile[:,:,i], cmap='gray')
    ax[1, i].imshow(image_cv2[:,:,i], cmap='gray')
for axi in ax.ravel():
    axi.set_axis_off()

plt.show()

The first row shows the four slices for the image read with tifffile. The second row shows the same, but read with opencv (cv2)

Opencv swaps channels 1 and 3. I learned it the hard way, and wanted to give you a heads-up! I hope you enjoyed reading.

Cheers :)