Daniel Drake wrote :
wxCover wrote:Well... I did it but I'm not sure it worked. I haven't received any confirmation or whatever. Should I retry?I just force-subscribed you. Let me know if there are any further problems sending or receiving the list mail.
I've received the confirmation. It should work now! thanks
Well, I've spoke too soon in my last mail! I hadn't tried to enroll and verify. The images I get looks nice BUT for the moment, I have problems with enrollment and verification. I think the main problem is that I'm not getting enough minutiae. Usually, for a fingerprint that looks really nice, I get something like 10~15 minutiae. And if the image is not perfect (very often), the number is less than 10.OK.Now, speaking about verification, I haven't manage to get it to match. The scores are always low. I think this is due to the poor number of minutiae. I don't know how to improve the situation. Is there some tweaking we can do with the algorithm which calculate minutiaes ?How low are the scores? Is there still a noticeable difference between scores of matching and non-matching finger scans?The default match score is 40, but the resultant score is dependant on the number of detected minutiae, so for small minutiae counts it is hard to reach such values.The driver can lower the match score threshold via the bz3_threshold attribute (e.g. aes4000 does this).You may also want to try enlarging the image with the enlarge_factor attribute. Again, aes4000 does this as the binarization results were horrible otherwise...The latest fprint_demo and libfprint from git will undoubtedly help you get a better understanding of the problems here. You can view the scanned image, the binarized image, and you can plot the detected minutiae on both.
I've installed both from git and tested all that stuff.There is a difference between the scores. I've changed bz3_threshold to 15. With this setting, it works better: I get many more matches. I haven't seen any wrong positives with that. But there are some false negatives. Maybe we could lower bz3_threshold a bit more? When I read another finger than the one enrolled, I never get scores higher than 7. The maximum score I got was 32. I've tried to change the enlarge factor but it didn't help (in fact, it may be worst=scores lower). Maybe because the binarization gives pretty good results (see the attached files)? In my last mail, I told the number of minutiae was very often lower than 10. In fact, now with "training", I'm getting much less "bad scan" errors...
In the attachments, you'll find 3 files: -two with the correct finger (the one enrolled) -one with another finger bz_xx is for the score and min_xx is the number of minutiae.In fact a big issue is to have a correct enrollment. I think in the future we'll need to read several times the same finger and see if there's no problem with it. Anyway, you may already know that!
Thanks, Daniel
<<inline: fingerprint_match_bz_23_min_12.png>>
<<inline: fingerprint_no_match_bz_10_min14.png>>
<<inline: fingerprint_no_match_other_finger_bz_0_min_10.png>>
_______________________________________________ fprint mailing list [email protected] http://lists.reactivated.net/mailman/listinfo/fprint
