Tuesday, February 16, 2010

Quantizing the colors in an image, using (server side) JavaScript



Top left: The original image. Top right: The image quantized to 4 bits of color information per channel. Lower left: 3 bits of color per channel. Lower right: 2 bits per channel.

It turns out to be surprisingly quick and easy to quantize the colors in an image to a smaller number of bits per channel than the standard 8 bits for red, 8 bits for green, and 8 bits for blue. All you have to do is loop over the pixels and AND them against the appropriate mask value. A mask value of 0xFFF0F0F0 discards the lower 4 bits' worth of color information from each channel, essentially leaving 4 bits, each, for red, green, and blue. A mask value of 0xFFE0E0E0 keeps just the top 3 bits in each channel, while a mask of 0xFFC0C0C0 retains just 2 bits of color per channel.

To obtain the images shown above, I ran the following script against them (using these various mask values) with the aid of the ImageMunger Java app that I gave code for earlier. The ImageMunger class simply opens an image of your choice (you supply the filepath as a command line argument) and runs the JavaScript file of your choice (a second command line argument), putting variables Image and Panel in scope at runtime. The Image variable is just a reference to the BufferedImage object, representing your image. The Panel variable is a reference to the JComponent in which ImageMunger draws your image.

MASK = 0xffc0c0c0; // 2 bits per channel
// 0xffe0e0e0 3 bits per channel
// 0xfff0f0f0 4 bits per channel

var w = Image.getWidth();
var h = Image.getHeight();
var pixels = Image.getRGB( 0,0,w,h,null,0,w );

for (i = 0, len = pixels.length; i < len; i++)
pixels[ i ] &= MASK;

Image.setRGB( 0,0,w,h,pixels,0,w );
Panel.updatePanel( );

The getRGB() method of BufferedImage fetches the pixels from your image as a giant one-dimensional array. The corresponding setImage() method replaces the pixels. The updatePanel() method of Panel (defined in ImageMunger.java) causes the JComponent to refresh.

Given that this is JavaScript and not Java, we shouldn't be surprised to find that performance isn't exactly breakneck. Still, at 110 pixels per millisecond, thoughput isn't terrible, either.

As you might expect, quantizing the color info makes the image easier to compress. The original image, in PNG form, occupies 185 Kbytes on disk. The 4-bit-per-channel version occupies just 61K; the 3-bit version, 38K; and the 2-bit version, a little over 23K.