Now let's vary the compression ratio and see what happens. The FBI's target bit rate is around 0.75 bits per pixel (bpp), which corresponds to a target compression ratio of 10.7 starting from an 8-bit image. This target bit rate is set via a ``knob'' on the WSQ algorithm, similar to the ``quality'' parameter in many JPEG implementations. In practice, the WSQ algorithm invariably yields a higher compression ratio (lower bit rate) than the target because of unpredictable amounts of lossless entropy coding gain, principally due to variable amounts of blank space in the images. E.g., fingerprints coded with WSQ at a target of 0.75 bpp will actually come in around 15:1 on average, although the exact compression ratio achieved will vary from image to image. For this reason, the WSQ target bit rate (as well as the JPEG ``quality'' parameter) should be thought of as setting the distortion level-rather than the precise compression ratio-for the compressed images.
Here's another sample image, 768 x 768 pixels (589824 bytes):
and here's a 4x zoom of the core:
Let's start by setting a target bit rate that's 20% higher than the FBI target of 0.75 bpp. A target of 0.9 bpp would correspond to a target compression ratio of 8.9, although the actual compressed WSQ file size of 47619 bytes corresponds to an achieved compression ratio of 12.4, attributable to the blank space surrounding the fingerprint. We then adjusted the JPEG quality parameter to try to match the compressed file sizes.
(Try to widen your web viewer enough to get the pictures side-by-side for comparison.)
WSQ image (left/top); file size 47619 bytes, compression ratio 12.4. JPEG image (right/bottom); file size 49658 bytes, compression ratio 11.9.
The WSQ image looks pretty good at this compression level, and while the JPEG image isn't bad, it's already showing blocking artifacts at less than 12:1 compression. (Since we couldn't match file sizes exactly, we were sporting and gave JPEG a 2 KB advantage, for an achieved compression ratio of 11.9.)
We should mention that these JPEG compressions were made using the ``standard'' JPEG quantization matrix. The Police Research Group of the UK Home Office has experimented with quantization matrices tuned specifically for fingerprint images, but the consensus between the FBI and the Home Office seems to be that JPEG blocking artifacts are unavoidable in fingerprints at compression ratios above about 10:1, even with customized quantization matrices.
Also, both the JPEG and WSQ compression examples used adaptive Huffman entropy coding.
Next, the same image is coded with WSQ at the current FBI target of 0.75 bpp and comes in at 39270 bytes (15.0:1); the closest we could get with JPEG was 40780 bytes (14.5:1).
WSQ image (left/top); file size 39270 bytes, compression ratio 15.0. JPEG image (right/bottom); file size 40780 bytes, compression ratio 14.5.
The JPEG blocking and loss of fine-scale features are becoming more objectionable.
Now let's see what happens if we overcompress this fingerprint by using a target bit rate 20% below 0.75 bpp. This experiment is of interest for a couple reasons. First, a system intended for real-world use must be robust enough to handle (gracefully!) the inevitable boo-boos resulting from misuse of equipment or lack of appropriate system maintenance or calibration. Second, such an experiment gives some indication of the potential for improving an algorithm for use at a higher compression ratio.
WSQ image (left/top); file size 30987 bytes, compression ratio 19.0. JPEG image (right/bottom); file size 30081 bytes, compression ratio 19.6.
The WSQ image is still serviceable at this level; the ridges are still distinguishable, although the fine details are fading. Improvements in the WSQ encoder, which are possible under the FBI standard, might produce acceptable image quality in the future at 0.6 bpp. The JPEG image, however, appears to be completely overwhelmed by blocking artifacts at this compression level.
Last updated 25 June 2002.