Tukukkk

I have an image size of 256x256. I will extract the image into patches with a size of 32x32, and the center of a patch will slide with window step of 20. If extracted patch size is smaller than 32x32. Then add zero padding to obtain 32x32. The code to extract patch is

import numpy as np

image = np.random.randn(256,256)

patch_H, patch_W = 32, 32

step_H, step_W = 20, 20



lst_H = np.arange(0, image.shape[0] + patch_H, step_H)

lst_W = np.arange(0, image.shape[1] + patch_W, step_W)

image_patches =

for i in range(len(lst_H)):

  for j in range(len(lst_W)):

    h= lst_H[i]

    w= lst_W[j]

    patch = image[h : h + patch_H, w : w + patch_W]

    # Zero padding

    if (patch.shape[0]!= patch_H or patch.shape[1]!= patch_W):

      patch = np.pad(patch,[(0, patch_H- patch.shape[0]), (0, patch_W- patch.shape[1])], mode='constant')

    patch =patch[None,:,:]

    image_patches.append(patch)

image_patches = np.vstack(image_patches) # (225, 32, 32)

Using the code above, I obtain 225 patches size of 32x32. From the image_patches, I want to generate a new image with the same size as the original image (256x256) and the values are summed up between overlapping patches.

The difficult is that I have to unpad the patches that are zero padding in the first step. Thus I cannot assign the image_rec[h : h + patch_H, w : w + patch_W]+=image_patches[num_patches,:,:] in bellow code if I did not unpad them. How could I solve the problem? In my opinion, I guess we have to save a dictionary to save the other keys such as zero flag, position...This is what I done

# Recover the image

image_rec = np.zeros (image.shape)

num_patches=0

for i in range(len(lst_H)):

  for j in range(len(lst_W)):

    h= lst_H[i]

    w= lst_W[j]

    image_rec[h : h + patch_H, w : w + patch_W]+=image_patches[num_patches,:,:]

    num_patches +=1