Uses chunks to upsample an image which reduces the required memory, especially helpful with very memory intensive models like EDSR.

super_resolution_with_chunks.py

import cv2
import numpy as np
import time

# Load the input image
input_image = cv2.imread('example.jpg')

# Define the size of the chunks
chunk_size = (200, 200)

# Create a DNN super-resolution object
sr = cv2.dnn_superres.DnnSuperResImpl_create()
sr.readModel('models/EDSR_x4.pb')
sr.setModel('edsr', 4)
sr.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
sr.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)

start = time.time()

# Compute the new image size based on the chunk size and scale factor
new_size = (input_image.shape[1] * 4, input_image.shape[0] * 4)

# Create an empty output image with the same number of channels as the input image
output_image = np.zeros((new_size[1], new_size[0], input_image.shape[2]), dtype=np.uint8)

# Loop over the image chunks
for y in range(0, input_image.shape[0], chunk_size[1]):
    for x in range(0, input_image.shape[1], chunk_size[0]):
        # Extract the chunk from the input image
        chunk = input_image[y:y+chunk_size[1], x:x+chunk_size[0], :]

        # Upsample the chunk
        upsampled_chunk = sr.upsample(chunk)

        # Insert the upscaled chunk into the output image
        output_image[y*4:(y+chunk_size[1])*4, x*4:(x+chunk_size[0])*4, :] = upsampled_chunk

end = time.time()
print("INFO: super resolution took {:.6f} seconds".format(end-start))

# Display the output image
cv2.imshow('Upscaled image', output_image)
cv2.imwrite('upscaled.jpg', output_image)
cv2.waitKey()
cv2.destroyAllWindows()