问题
I have an image of shape (466,394,1) which I want to split into 7x7 patches.
image = tf.placeholder(dtype=tf.float32, shape=[1, 466, 394, 1])
Using
image_patches = tf.extract_image_patches(image, [1, 7, 7, 1], [1, 7, 7, 1], [1, 1, 1, 1], 'VALID')
# shape (1, 66, 56, 49)
image_patches_reshaped = tf.reshape(image_patches, [-1, 7, 7, 1])
# shape (3696, 7, 7, 1)
unfortunately does not work in practice as image_patches_reshaped mixes up the pixel order (if you view images_patches_reshaped you will only see noise).
So my new approach was to use tf.split:
image_hsplits = tf.split(1, 4, image_resized)
# [<tf.Tensor 'split_255:0' shape=(462, 7, 1) dtype=float32>,...]
image_patches = []
for split in image_hsplits:
image_patches.extend(tf.split(0, 66, split))
image_patches
# [<tf.Tensor 'split_317:0' shape=(7, 7, 1) dtype=float32>, ...]
this indeed preserves the image pixel order unfortunately it creates a lot of OPs which is not very good.
How do I split an image into smaller patches with less OPs?
Update1:
I ported the answer of this question for numpy to tensorflow:
def image_to_patches(image, image_height, image_width, patch_height, patch_width):
height = math.ceil(image_height/patch_height)*patch_height
width = math.ceil(image_width/patch_width)*patch_width
image_resized = tf.squeeze(tf.image.resize_image_with_crop_or_pad(image, height, width))
image_reshaped = tf.reshape(image_resized, [height // patch_height, patch_height, -1, patch_width])
image_transposed = tf.transpose(image_reshaped, [0, 2, 1, 3])
return tf.reshape(image_transposed, [-1, patch_height, patch_width, 1])
but I think there is still room for improvement.
Update2:
This will convert patches back to the original image.
def patches_to_image(patches, image_height, image_width, patch_height, patch_width):
height = math.ceil(image_height/patch_height)*patch_height
width = math.ceil(image_width/patch_width)*patch_width
image_reshaped = tf.reshape(tf.squeeze(patches), [height // patch_height, width // patch_width, patch_height, patch_width])
image_transposed = tf.transpose(image_reshaped, [0, 2, 1, 3])
image_resized = tf.reshape(image_transposed, [height, width, 1])
return tf.image.resize_image_with_crop_or_pad(image_resized, image_height, image_width)
回答1:
I think your issue is somewhere else. I wrote the following code snippet (using a smaller 14x14 image so that I could hand-check all the values), and confirmed that your initial code did the correct operations:
import tensorflow as tf
import numpy as np
IMAGE_SIZE = [1, 14, 14, 1]
PATCH_SIZE = [1, 7, 7, 1]
input_image = np.reshape(np.array(xrange(14*14)), IMAGE_SIZE)
image = tf.placeholder(dtype=tf.int32, shape=IMAGE_SIZE)
image_patches = tf.extract_image_patches(
image, PATCH_SIZE, PATCH_SIZE, [1, 1, 1, 1], 'VALID')
image_patches_reshaped = tf.reshape(image_patches, [-1, 7, 7, 1])
sess = tf.Session()
(output, output_reshaped) = sess.run(
(image_patches, image_patches_reshaped),
feed_dict={image: input_image})
print "Output (shape: %s):" % (output.shape,)
print output
print "Reshaped (shape: %s):" % (output_reshaped.shape,)
print output_reshaped
The output was:
python resize.py
Output (shape: (1, 2, 2, 49)):
[[[[ 0 1 2 3 4 5 6 14 15 16 17 18 19 20 28 29 30 31
32 33 34 42 43 44 45 46 47 48 56 57 58 59 60 61 62 70
71 72 73 74 75 76 84 85 86 87 88 89 90]
[ 7 8 9 10 11 12 13 21 22 23 24 25 26 27 35 36 37 38
39 40 41 49 50 51 52 53 54 55 63 64 65 66 67 68 69 77
78 79 80 81 82 83 91 92 93 94 95 96 97]]
[[ 98 99 100 101 102 103 104 112 113 114 115 116 117 118 126 127 128 129
130 131 132 140 141 142 143 144 145 146 154 155 156 157 158 159 160 168
169 170 171 172 173 174 182 183 184 185 186 187 188]
[105 106 107 108 109 110 111 119 120 121 122 123 124 125 133 134 135 136
137 138 139 147 148 149 150 151 152 153 161 162 163 164 165 166 167 175
176 177 178 179 180 181 189 190 191 192 193 194 195]]]]
Reshaped (shape: (4, 7, 7, 1)):
[[[[ 0]
[ 1]
[ 2]
[ 3]
[ 4]
[ 5]
[ 6]]
[[ 14]
[ 15]
[ 16]
[ 17]
[ 18]
[ 19]
[ 20]]
[[ 28]
[ 29]
[ 30]
[ 31]
[ 32]
[ 33]
[ 34]]
[[ 42]
[ 43]
[ 44]
[ 45]
[ 46]
[ 47]
[ 48]]
[[ 56]
[ 57]
[ 58]
[ 59]
[ 60]
[ 61]
[ 62]]
[[ 70]
[ 71]
[ 72]
[ 73]
[ 74]
[ 75]
[ 76]]
[[ 84]
[ 85]
[ 86]
[ 87]
[ 88]
[ 89]
[ 90]]]
[[[ 7]
[ 8]
[ 9]
[ 10]
[ 11]
[ 12]
[ 13]]
[[ 21]
[ 22]
[ 23]
[ 24]
[ 25]
[ 26]
[ 27]]
[[ 35]
[ 36]
[ 37]
[ 38]
[ 39]
[ 40]
[ 41]]
[[ 49]
[ 50]
[ 51]
[ 52]
[ 53]
[ 54]
[ 55]]
[[ 63]
[ 64]
[ 65]
[ 66]
[ 67]
[ 68]
[ 69]]
[[ 77]
[ 78]
[ 79]
[ 80]
[ 81]
[ 82]
[ 83]]
[[ 91]
[ 92]
[ 93]
[ 94]
[ 95]
[ 96]
[ 97]]]
[[[ 98]
[ 99]
[100]
[101]
[102]
[103]
[104]]
[[112]
[113]
[114]
[115]
[116]
[117]
[118]]
[[126]
[127]
[128]
[129]
[130]
[131]
[132]]
[[140]
[141]
[142]
[143]
[144]
[145]
[146]]
[[154]
[155]
[156]
[157]
[158]
[159]
[160]]
[[168]
[169]
[170]
[171]
[172]
[173]
[174]]
[[182]
[183]
[184]
[185]
[186]
[187]
[188]]]
[[[105]
[106]
[107]
[108]
[109]
[110]
[111]]
[[119]
[120]
[121]
[122]
[123]
[124]
[125]]
[[133]
[134]
[135]
[136]
[137]
[138]
[139]]
[[147]
[148]
[149]
[150]
[151]
[152]
[153]]
[[161]
[162]
[163]
[164]
[165]
[166]
[167]]
[[175]
[176]
[177]
[178]
[179]
[180]
[181]]
[[189]
[190]
[191]
[192]
[193]
[194]
[195]]]]
Based on the reshaped output, you can see it is a 4x7x7x1 with values for the first patch as: [0-7),[14-21), [28-35), [42-49), [56-63), [70-77), and [84-91), which corresponds to the upper left 7x7 grid.
Perhaps you can explain a bit further what's going on when it doesn't work correctly?
来源：https://stackoverflow.com/questions/41564321/split-image-tensor-into-small-patches