Error Converting PIL B&W images to Numpy Arrays

Question

I am getting weird errors when I try to convert a black and white PIL image to a numpy array. An example of the code I am working with is below.

    if image.mode != '1':
        image = image.convert('1') #convert to B&W
    data = np.array(image) #Have also tried np.asarray(image)
    n_lines = data.shape[0] #number of raster passes
    line_range = range(data.shape[1])
    for l in range(n_lines):
        # process one horizontal line of the image
        line = data[l]
        for n in line_range:
            if line[n] == 1:
                write_line_to(xl, z+scale*n, speed) #conversion to other program code
            elif line[n] == 0:
                run_to(xl, z+scale*n) #conversion to other program code

I have tried this using both array and asarray for the conversion, and gotten different errors. If I use array, then the data I get out is nothing like what I put in. It looks like several very shrunken partial images side by side, with the remainder of the image space filled in in black. If I use asarray, then the entirety of python crashes during the raster step (on a random line). If I work with a greyscale image ('L'), then neither of these errors occurs for either array or asarray.

Does anyone know what I am doing wrong? Is there something odd about the way PIL encodes B&W images, or something special I need to pass numpy to make it convert properly?

A subsidiary issue is that we just found is that the conversion appears to be dithering, where we want to have solid lines. — Elliot, May 04 '10 at 14:35
In that case (dithering not desirable) it probably makes the most sense to keep the images as grayscale (mode 'L' / dtype np.uint8) and threshold them using numpy. E.g. "data = np.array(im); data = data > 127" should work fine for grayscale images, and avoid dithering altogether. You can also threshold things in PIL, but if you're converting to numpy arrays anyway, it's easier to do it with numpy. Just my thoughts, anyway... Good luck! — Joe Kington, May 04 '10 at 19:03

score 7 · Accepted Answer · answered May 04 '10 at 17:33

I believe you've found a bug in PIL! (or possibly in numpy, but I'd wager it's on the PIL side of things...)

@c's answer above gives one workaround (use im.getdata()), though I'm not sure why numpy.asarry(image) is segfaulting for him... (Old version of PIL and/or numpy, maybe?) It works for me, but produces gibberish on 1-bit PIL images (and works for everything else, I use it frequently!).

Another workaround is to convert the BW image back to grayscale (mode 'L') before converting to a numpy array.

Converting the BW image back to grayscale before converting to a numpy array seems to be faster, if speed is important.

In [35]: %timeit np.array(im_bw.convert('L')).astype(np.uint8)
10000 loops, best of 3: 28 us per loop

In [36]: %timeit np.reshape(im_bw.getdata(), im_bw.size)
10000 loops, best of 3: 57.3 us per loop

On a seperate note, if you're modifying the array contents in-place, be sure to use numpy.array instead of numpy.asarray, as the latter will create an array from the PIL image instance without copying memory, thus returning a read-only array. Just mentioning this because I'm using asarray() below...

Here's a standalone example which confirms the bug...

import numpy as np
import Image

x = np.arange(256, dtype=np.uint8).reshape((16,16))
print 'Created array'
print x

im = Image.fromarray(x)
print 'Vales in grayscale PIL image using numpy.asarray <-- Works as expected'
print np.asarray(im)

print 'Converted to BW PIL image...'
im_bw = im.convert('1')

print 'Values in BW PIL image, using Image.getdata() <-- Works as expected'
print '  (Not a simple threshold due to dithering...)'
# Dividing by 255 to make the comparison easier
print np.reshape(im_bw.getdata(), (16, 16)) / 255 

print 'Values in BW PIL image using numpy.asarray() <-- Unexpected!'
print '   (Same occurs when using numpy.array() to copy and convert.)'
print np.asarray(im_bw).astype(np.uint8) 

print 'Workaround, convert back to type "L" before array conversion'
print np.array(im_bw.convert('L')).astype(np.uint8) / 255

Which outputs:

Created array
[[  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15]
 [ 16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31]
 [ 32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47]
 [ 48  49  50  51  52  53  54  55  56  57  58  59  60  61  62  63]
 [ 64  65  66  67  68  69  70  71  72  73  74  75  76  77  78  79]
 [ 80  81  82  83  84  85  86  87  88  89  90  91  92  93  94  95]
 [ 96  97  98  99 100 101 102 103 104 105 106 107 108 109 110 111]
 [112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127]
 [128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143]
 [144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159]
 [160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175]
 [176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191]
 [192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207]
 [208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223]
 [224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239]
 [240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255]]

Vales in grayscale PIL image using numpy.asarray <-- Works as expected
[[  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15]
 [ 16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31]
 [ 32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47]
 [ 48  49  50  51  52  53  54  55  56  57  58  59  60  61  62  63]
 [ 64  65  66  67  68  69  70  71  72  73  74  75  76  77  78  79]
 [ 80  81  82  83  84  85  86  87  88  89  90  91  92  93  94  95]
 [ 96  97  98  99 100 101 102 103 104 105 106 107 108 109 110 111]
 [112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127]
 [128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143]
 [144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159]
 [160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175]
 [176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191]
 [192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207]
 [208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223]
 [224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239]
 [240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255]]

Converted to BW PIL image...

Values in BW PIL image, using Image.getdata() <-- Works as expected
  (Not a simple threshold due to dithering...)
[[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0]
 [0 0 1 0 1 0 1 0 1 0 0 0 1 0 0 0]
 [0 1 0 0 0 0 0 0 0 0 1 0 0 1 0 1]
 [0 0 0 1 0 1 0 1 0 1 0 1 0 0 0 0]
 [1 0 1 0 1 0 1 0 1 0 0 0 1 1 0 1]
 [0 1 0 1 0 0 1 0 0 1 1 0 1 0 1 0]
 [1 0 1 0 1 0 1 1 0 1 0 1 0 1 0 1]
 [0 1 0 1 0 1 0 1 0 1 1 0 1 1 0 1]
 [1 1 0 1 1 1 1 0 1 1 0 1 1 0 1 1]
 [1 0 1 1 0 1 0 1 1 0 1 1 0 1 1 0]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 0 1 0 1 1 0 1 1 0 1 1 1 0 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]]

Values in BW PIL image using numpy.asarray() <-- Unexpected!
   (Same occurs when using numpy.array() to copy and convert.)
[[0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [0 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]]

Workaround, convert back to type "L" before array conversion
[[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0]
 [0 0 1 0 1 0 1 0 1 0 0 0 1 0 0 0]
 [0 1 0 0 0 0 0 0 0 0 1 0 0 1 0 1]
 [0 0 0 1 0 1 0 1 0 1 0 1 0 0 0 0]
 [1 0 1 0 1 0 1 0 1 0 0 0 1 1 0 1]
 [0 1 0 1 0 0 1 0 0 1 1 0 1 0 1 0]
 [1 0 1 0 1 0 1 1 0 1 0 1 0 1 0 1]
 [0 1 0 1 0 1 0 1 0 1 1 0 1 1 0 1]
 [1 1 0 1 1 1 1 0 1 1 0 1 1 0 1 1]
 [1 0 1 1 0 1 0 1 1 0 1 1 0 1 1 0]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 0 1 0 1 1 0 1 1 0 1 1 1 0 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]]

I don't think this bug has gone away. I'm still getting `print(np.array(Image.new('1', (4, 2), color=1)))` -> `[[ True True False False] [False False False False]]` on my raspberry pi with `Image.VERSION` 1.1.7 and `np.__version__` 1.13.3. But interestingly, on Mac OS X I get the correct result: `[[ True True True True] [ True True True True]]` with the same versions of both. — Bill, Nov 05 '17 at 04:42

score 3 · Answer 2 · answered May 04 '10 at 00:59

Not sure about this line:

data = numpy.array(image)

In fact, that gives me a segfault. But I just tried the following, and it works fine:

import numpy
import Image

im = Image.open("some_photo.jpg")
im = im.convert("1")

pixels = im.getdata() # returns 1D list of pixels
n = len(pixels)
data = numpy.reshape(pixels, im.size) # turn into 2D numpy array

for row in data:
    # do your processing
    pass

# Check that the numpy array's data is good
im2 = Image.new("1", im.size)
im2.putdata(numpy.reshape(data, [n, 1]))
im2.show()

Something is weird about the b&W conversion. It is not properly registering edges. So instead of a logo, I am getting a big circle. — Elliot, May 04 '10 at 15:37

score 0 · Answer 3 · answered Jul 05 '21 at 10:18

0

What's your numpy version? I found that after downgrading of numpy from 1.21 to 1.20, it worked.

pip install numpy==1.20

answered Jul 05 '21 at 10:18

Jaeyoon Jeong

569
4
8

Error Converting PIL B&W images to Numpy Arrays

3 Answers3

Linked