Compression filters are a way of transforming the image data (without loss of information) so that it will compress better. Each row in the image can have one of five filter types associated with it; choosing which of the five to use for each row is almost more of a black art than a science. Nevertheless, at least one reasonably good algorithm is not only known but is also described in the PNG specification and is implemented in freely available software. Other algorithms are likely to perform even better, but so far this has not been an active area of research.

By way of example--admittedly an extreme case--a 512 × 32,768 image containing all 16,777,216 possible 24-bit colors compressed over 300 times better with filtering than without. The uncompressed image was 48 MB in size; the compressed but unfiltered version was around 36 MB; but the filtered version (using the ``reasonably good algorithm'' referred to earlier) was only 115,989 bytes (0.1 MB). And a version created by trying multiple filtering approaches was a mere 91,569 bytes, for a total compression ratio of 550:1 and an improvement over the unfiltered version of more than 400 times. Keep in mind that we're talking about completely lossless compression here. Yow.

Filtering is also described in more detail in Chapter 9, "Compression and Filtering".