Matching result differences in Java and C# implementations

[gustingonzalez]

Hi,

I'm currently working with fingerprint images in the WSQ format; but I see the WSQ support is not yet available in the C# version of SourceAFIS.

To work around this, I tried using the Java version, which supports WSQ images. While this approach worked, I noticed a slight difference in matching results compared to earlier tests with the C# version using PNG images. Specifically, the Java version shown lower precision, including a false positive match that did not occur with the C# version.

Given this, could you recommend a reliable library for decoding WSQ images so that I can preprocess them and continue using the .NET version? I tried NBIS (using dwsq), but was unable to export the WSQ images correctly. On the other hand, if I preprocess the images by converting it first from WSQ, would the minutiae extraction perform worse compared to working directly with WSQ images?

Thanks in advance!

[robertvazan]

The algorithm should be identical in Java/C# and in WSQ/PNG. The difference is probably in WSQ compression artifacts and perhaps in color space. No recommendations for WSQ library, unfortunately.

[gustingonzalez]

Thank you for your quick response, @robertvazan.

This might belong to a different issue, but I would like to share some results from a quick experiment comparing the Java and C# versions of SourceAFIS using PNG images (no WSQ involved). Here's a cropped version of the comparison table:

		SourceAFIS C#		SourceAFIS Java
Probe	Expected	Retrieved	Score	Retrieved	Score
8293-1-fp3.png	8293-1-fp1.png	fvc2002/DB2_B/109_8.tif	20.11	8293-1-fp1.png	61.08
8293-2-fp3.png	8293-1-fp1.png	fvc2002/DB4_B/108_4.tif	19.72	8293-1-fp1.png	45.4
2839-1-fp3.png	2839-fp1.png	2839-fp1.png	165.18	fvc2002/DB1_B/105_8.tif	26.71
2839-2-fp3.png	2839-fp1.png	2839-fp1.png	191.11	fvc2002/DB3_B/106_8.tif	42.66

The results show that, in the first two cases, the Java version retrieves the correct match with a "good" score, while the C# version does not. In the last two cases, it’s the opposite: the C# version retrieves the expected match with a significantly "good" score, while the Java version does not—and in fact returns a false positive, with a score above the threshold of 40. Interestingly, for this last case, the same false positive appears in the C# ranking as well (not shown in the table), but with a lower score (28.38), below the recommended threshold.

If the score algorithm is indeed the same across both versions, I think the differences might be due to how minutiae are extracted or how image preprocessing is handled. I didn't do much research into why this happens, but it's perhaps something worth considering.

[robertvazan]

Hmm, looking at the output of SourceAFIS CLI, overall accuracy is the same even though there are minor differences, which I believe are due to one unresolved consistency bug. The CLI however uses raw grayscale images that were converted ahead of the time from source images using one common image decoder to ensure consistency. So the discrepancies you are seeing could be either due to PNG decoder differences or due to the consistency bug, but overall accuracy should be nearly the same.

[gustingonzalez]

Ok, got it, @robertvazan. I'll consider image preprocessing in light of the differences in results between both implementations.

On the other hand, I found a way to convert WSQ images using this package, but the recommended method by the author didn’t work for me. Instead, I used the following snippet, that works well:

import _wsq
fingerprint_wsq_data = load_wsq(...)
buf, w, h, ppi = _wsq.decompress(fingerprint_wsq_data)
img = Image.frombytes('L', (w, h), buf)

# Add dpi info.
if ppi > 0:
    img.save("output.tif", dpi=(ppi, ppi))
else:
    img.save("output.tif")

Finally, my issue with the NBIS utilities was that I didn't know how to use them to convert a WSQ file into a "readable" image (PNG, BMP, etc.). That said, I believe something similar to the above snippet could be done by using the dwsq utility and, then, processing (for example, by using PIL or OpenCV) the binary data and the info of the generated .ncm as header.