PurePNG Code Examples

This section discusses some example Python programs that use the png module for reading and writing PNG files.

Writing

The basic strategy is to create a Writer object (instance of png.Writer) and then call its png.write() method with an open (binary) file, and the pixel data. The Writer object encapsulates all the information about the PNG file: image size, colour, bit depth, and so on.

A Ramp

Create a one row image, that has all grey values from 0 to 255. This is a bit like Netpbm’s pgmramp.

import png
f = open('ramp.png', 'wb')      # binary mode is important
w = png.Writer(255, 1, greyscale=True)
w.write(f, [range(256)])
f.close()

Note that our single row, generated by range(256), must itself be enclosed in a list. That’s because the png.write() method expects a list of rows.

From now on import png will not be mentioned.

A Little Message

A list of strings holds a graphic in ASCII graphic form. We convert it to a list of integer lists (the required form for the write() method), and write it out as a black-and-white PNG (bilevel greyscale).

s = ['110010010011',
     '101011010100',
     '110010110101',
     '100010010011']
s = map(lambda x: map(int, x), s)

f = open('png.png', 'wb')
w = png.Writer(len(s[0]), len(s), greyscale=True, bitdepth=1)
w.write(f, s)
f.close()

Note how we use len(s[0]) (the length of the first row) for the x argument and len(s) (the number of rows) for the y argument.

A Palette

The previous example, “a little message”, can be converted to colour simply by creating a PNG file with a palette. The only difference is that a palette argument is passed to the write() method instead of greyscale=True:

# Assume f and s have been set up as per previous example
palette=[(0x55,0x55,0x55), (0xff,0x99,0x99)]
w = png.Writer(len(s[0]), len(s), palette=palette, bitdepth=1)
w.write(f, s)

Note that the palette consists of two entries (the bit depth is 1 so there are only 2 possible colours). Each entry is an RGB triple. If we wanted transparency then we can use RGBA 4-tuples for each palette entry.

Colour

For colour images the input rows are generally 3 times as long as for greyscale, because there are 3 channels, RGB, instead of just one, grey. Below, the p literal has 2 rows of 9 values (3 RGB pixels per row). The spaces are just for your benefit, to mark out the separate pixels; they have no meaning in the code.

p = [(255,0,0, 0,255,0, 0,0,255),
     (128,0,0, 0,128,0, 0,0,128)]
f = open('swatch.png', 'wb')
w = png.Writer(3, 2)
w.write(f, p) ; f.close()

More Colour

A further colour example illustrates some of the manoeuvres you have to perform in Python to get the pixel data in the right format.

Say we want to produce a PNG image with 1 row of 8 pixels, with all the colours from a 3-bit colour system (with 1-bit for each channel; such systems were common on 8-bit micros from the 1980s).

We produce all possible 3-bit numbers:

>>> range(8)
[0, 1, 2, 3, 4, 5, 6, 7]

We can convert each number into an RGB triple by assigning bit 0 to blue, bit 1 to red, bit 2 to green (the convention used by a certain 8-bit micro):

>>> map(lambda x: (bool(x&2), bool(x&4), bool(x&1)), _)
[(False, False, False), (False, False, True), (True, False, False),
(True, False, True), (False, True, False), (False, True, True), (True,
True, False), (True, True, True)]

(later on we will convert False into 0, and True into 255, so don’t worry about that just yet). Here we have each pixel as a tuple. We want to flatten the pixels so that we have just one row. In other words instead of [(R,G,B), (R,G,B), ...] we want [R,G,B,R,G,B,...]. It turns out that itertools.chain(*...) is just what we need:

>>> list(itertools.chain(*_))
[False, False, False, False, False, True, True, False, False, True,
False, True, False, True, False, False, True, True, True, True, False,
True, True, True]

Note that the list is not necessary, we can usually use the iterator directly instead. I just used list here so we can see the result.

Now to convert False to 0 and True to 255 we can multiply by 255 (Python use’s Iverson’s convention, so False==0, True==1). We could do that with map(lambda x:255*x, _). Or, we could use a “magic” bound method:

>>> map((255).__mul__, _)
[0, 0, 0, 0, 0, 255, 255, 0, 0, 255, 0, 255, 0, 255, 0, 0, 255, 255,
255, 255, 0, 255, 255, 255]

Now we write the PNG file out:

>>> p=_
>>> f=open('speccy.png', 'wb')
>>> w.write(f, [p]) ; f.close()

Reading

The basic strategy is to create a Reader object (a png.Reader instance), then call its png.read() method to extract the size, and pixel data.

PngSuite

The Reader() constructor can take either a filename, a file-like object, or a sequence of bytes directly. Here we use urllib to download a PNG file from the internet.

>>> r=png.Reader(file=urllib.urlopen('http://www.schaik.com/pngsuite/basn0g02.png'))
>>> r.read()
(32, 32, <itertools.imap object at 0x10b7eb0>, {'greyscale': True,
'alpha': False, 'interlace': 0, 'bitdepth': 2, 'gamma': 1.0})

The png.read() method returns a 4-tuple. Note that the pixels are returned as an iterator (not always, and the interface doesn’t guarantee it; the returned value might be an iterator or a sequence).

>>> l=list(_[2])
>>> l[0]
array('B', [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0,
1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3])

We have extracted the top row of the image. Note that the row itself is an array (see module array), but in general any suitable sequence type may be returned by read(). The values in the row are all integers less than 4, because the image has a bit depth of 2.

NumPy

NumPy is a package for scientific computing with Python. It is not part of a standard Python installation, it is downloaded and installed separately if needed. Numpy’s array manipulation facilities make it good for doing certain type of image processing, and scientific users of NumPy may wish to output PNG files for visualisation.

PyPNG does not have any direct integration with NumPy, but the basic data format used by PyPNG, an iterator over rows, is fairly easy to get into two- or three-dimensional NumPy arrays.

The code in this section is extracted from exnumpy.py, which is a complete runnable example in the code/ subdirectory of the source distribution. Code was originally written by Mel Raab, but has been hacked around since then.

PNG to NumPy array (reading)

The best thing to do (I think) is to convert each PyPNG row to a 1-dimensional numpy array, then stack all of those arrays together to make a 2-dimensional array. A number of features make this surprising compact. Say pngdata is the row iterator returned from png.Reader.asDirect(). The following code will slurp it into a 2-dimensional numpy array:

image_2d = numpy.vstack(itertools.imap(numpy.uint16, pngdata))

Note that the use of numpy.uint16, above, means that an array with data type numpy.uint16 is created which is suitable for bit depth 16 images. Replace numpy.uint16 with numpy.uint8 to create an array with a byte data type (suitable for bit depths up to 8).

Reshaping

For some operations it’s easier to have the image data in a 3-dimensional array. This plays to NumPy’s strengths:

image_3d = numpy.reshape(image_2d,
                         (row_count,column_count,plane_count))

NumPy array to PNG (writing)

Reshape your NumPy data into a 2-dimensional array, then use the fact that a NumPy array is an iterator over its rows:

    pngWriter.write(pngfile,
                    numpy.reshape(image_3d, (-1, column_count*plane_count)))

Currently (writing on 2009-04-16) this generates a warning; this warning appears to be a bug/limitation in NumPy, but it is harmless.