Reconstructing an image after using extract_image_patches

Question

I have an autoencoder that takes an image as an input and produces a new image as an output.

The input image (1x1024x1024x3) is split into patches (1024x32x32x3) bef

Answer 1

This code works for your specific case, as well as for cases when the images are square, with a square kernel and the image size is divisible by the kernel size.

I did not test it for other cases.

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt


size = 1024
k_size = 32
axes_1_2_size = int(np.sqrt((size * size) / (k_size * k_size)))

# Define a placeholder for image (or load it directly if you prefer) 
img = tf.placeholder(tf.int32, shape=(1, size, size, 3))

# Extract patches
patches = tf.image.extract_image_patches(img, ksizes=[1, k_size, k_size, 1], 
                                         strides=[1, k_size, k_size, 1], 
                                         rates=[1, 1, 1, 1], padding='VALID')

# Reconstruct the image back from the patches
# First separate out the channel dimension
reconstruct = tf.reshape(patches, (1, axes_1_2_size, axes_1_2_size, k_size, k_size, 3)) 
# Tranpose the axes (I got this axes tuple for transpose via experimentation)
reconstruct = tf.transpose(reconstruct, (0, 1, 3, 2, 4, 5))
# Reshape back
reconstruct = tf.reshape(reconstruct, (size, size, 3))

im_arr = # load image with shape (size, size, 3)

# Run the operations
with tf.Session() as sess:
    ps, r = sess.run([patches, reconstruct], feed_dict={img:[im_arr]})

# Plot the reconstructed image to verify
plt.imshow(r)