将图像的补丁拼接在一起

Question

您好，我有一批图像，我需要将其分成不重叠的补丁，并通过 softmax 函数发送每个补丁，然后重建原始图像。我可以按如下方式制作补丁：

@tf.function
def grid_img(img,patch_size=(256, 256), padding="VALID"):
    p_height, p_width = patch_size
    batch_size, height, width, n_filters = img.shape
    p = tf.image.extract_patches(images=img,
                       sizes=[1,p_height, p_width, 1],
                       strides=[1,p_height, p_width, 1],
                       rates=[1, 1, 1, 1],
                       padding=padding)
    new_shape = list(p.shape[1:-1])+[p_height, p_width, n_filters]
    p = tf.keras.layers.Reshape(new_shape)(p)
    return p

但是我不知道如何批量重建原始图像。对原始批次的简单重塑不起作用。数据不会以正确的方式排列。我将不胜感激任何帮助。谢谢

Answer 1

IIUC，你应该能够简单地使用tf.reshape从批量补丁中重建原始图像：

import tensorflow as tf

samples = 5
images = tf.random.normal((samples, 256, 256, 3))

@tf.function
def grid(images):
  img_shape = tf.shape(images)
  batch_size, height, width, n_filters = img_shape[0], img_shape[1], img_shape[2], img_shape[3]

  patches = tf.image.extract_patches(images=images,
                                      sizes=[1, 32, 32, 1],
                                      strides=[1, 32, 32, 1],
                                      rates=[1, 1, 1, 1],
                                      padding='VALID')
  return tf.reshape(tf.nn.softmax(patches), (batch_size, height, width, n_filters))
  
patches = grid(images)
print(patches.shape)
# (5, 256, 256, 3)

更新一：如果你想以正确的顺序重建图像，你可以计算 tf.image.extract_patches 的梯度，如代码 snippet 所示。这是一个例子：

import tensorflow as tf
import matplotlib.pyplot as plt
import pathlib

@tf.function
def grid(images):
  img_shape = tf.shape(images)
  patches = tf.image.extract_patches(images=images,
                                      sizes=[1, 64, 64, 1],
                                      strides=[1, 64, 64, 1],
                                      rates=[1, 1, 1, 1],
                                      padding='VALID')
  return patches

@tf.function
def extract_patches_inverse(shape, patches):
    _x = tf.zeros(shape)
    _y = grid(_x)
    grad = tf.gradients(_y, _x)[0]
    return tf.gradients(_y, _x, grad_ys=patches)[0] / grad


dataset_url = "https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz"
data_dir = tf.keras.utils.get_file('flower_photos', origin=dataset_url, untar=True)
data_dir = pathlib.Path(data_dir)
batch_size = 32

train_ds = tf.keras.utils.image_dataset_from_directory(
  data_dir,
  seed=123,
  image_size=(512, 512),
  batch_size = batch_size, 
  shuffle= False)

images, _ = next(iter(train_ds.skip(1).take(2)))
patches = grid(images)

shape = (batch_size, 512, 512, 3)
images_reconstructed = extract_patches_inverse(shape, patches)

plt.figure()
f, axarr = plt.subplots(1,2) 
axarr[0].imshow(images[0]/ 255)
axarr[1].imshow(images_reconstructed[0] / 255)

Answer 2

我想到的一个肮脏的工作是在转换后跟踪单元格的位置。不像@alonetogether Answer 那样优雅，但仍然可能有助于分享。

import numpy as np 
import tensorflow as tf

@tf.function
def grid(images, grid_size=(32, 32)):
    grid_height, grid_width = grid_size
    patches = tf.image.extract_patches(images=images,
                                      sizes=[1, grid_height, grid_width, 1],
                                      strides=[1, grid_height, grid_width, 1],
                                      rates=[1, 1, 1, 1],
                                      padding='VALID')
    return patches

batch_size, height, width, n_filters = shape = (5, 256, 256, 1)
indices = tf.range(batch_size * height * width * n_filters)
images = tf.reshape(indices, (batch_size, height, width, n_filters ))

patches = grid(images)
transfered_indices = tf.reshape(patches, shape=[-1])
tracked_indices = tf.argsort(transfered_indices) # Indices after transformation, Save this 


images = tf.random.normal(shape)

patches = grid(images)

flatten_patches = tf.reshape(patches, shape=[-1])

reconstructed = tf.reshape(tf.gather(flatten_patches, tracked_indices), shape)

np.alltrue(reconstructed==images) # True

将图像的补丁拼接在一起

Stitching patches of Image together

python

tensorflow

image-processing

tensorflow2.0