Suppose that we have a fingerprint database consisting of 1000 images from 100 persons (10 images per person), and it is required to design a system to identify the person from his fingerprint image. I know that there is a difference between identification and classification, but can we treat the identification problem as a classification problem by setting the number of classes to be equal to the number of persons (100 in this case)? I know that the identification accuracy is the metric used for assessing the performance of identification systems. For the classification problem, other metrics such as precision, recall and F1 score can be used, in addition to accuracy. Can we use these metrics with the identification problem or not? Is the accuracy the only metric that be used with the identification problem?
Dear Asmaa Diab
The risks of using biometrics fall into a few categories, including data and network hacking, rapidly evolving fraud capabilities, biometric enrollment security, familiar fraud (that is, caused by a family member or friend), spoofed sensors, and sensor inaccuracy. One of the greatest risks is data security.
Kind Regards
Qamar Ul Islam
Asmaa Diab ,
As you already stated there is a difference between classification and identification. But lets make it clearer so we can discuss the subject. In identification you are trying to pick a unique signature and differentiate it among N other possible signatures. That is you are trying to find the uniqueness of that signature (fingerprints) which in this case is the minutiae.
The classification problem is different in that you are trying to find a class on which to group together common elements.
If you see classification as a way to solve your problem of identification you would need to do at least two different operations. Try to reduce intra differences of the class (the individual) while maximising inter differences of classes (between individuals). The problem with this goal is how to achieve this with your dataset. That is, how do you guarantee there are no similar features that belong to different individuals?
The only way I would approach this is to train , say, a neural network to overfit the training so that it memorises the training set.This would make the system brittle, since distortions are common in fingerprint acquisition. At this point I would not see the advantage of using the NN(or similar ML classification algorithm) over traditional correlation, minutiae location and point pattern based approaches.
As a side note, the classification approach is used mainly for indexing of fingerprints and when you are using multi modal biometrics where it is useful and holds advantages over traditional methods.
Hope this helps
Arturo Geigel
https://www.mathworks.com/matlabcentral/fileexchange/60900-multi-class-confusion-matrix?s_tid=prof_contriblnk&fbclid=IwAR10C-A_4MQw4Cd0__2FXxymrV-xkXv02cxPP_6KJeiMQuCWEG5kOGvkENk
The above link contains performance assessment metrics for the classification problems, such as precision, recall, f1 score and accuracy. My question is: Can I use these metrics with the identification problem?? In all papers I read, I found that the only metric used is the accuracy. Can we use precision,recall and f1 score with the identification problem? Or they are only used with classification assessment?
Asmaa Diab ,
You can always add and/or modify existing metrics to further quantify your work. I do not see why not include some of those measurements as part of your work .
Note that, though these have to be contextualised. For example, I would use precision and recall not as part of the matching process but as a measure of the retrieval system of candidates on which to do the matching. This would be a different area of research.
Regards
Arturo Geigel Thanks a lot for your help. I would be grateful if you could explain the last paragraph of your answer with a simple example.
Asmaa Diab ,
Let's take the example and beginning explanation from:
https://towardsdatascience.com/beyond-accuracy-precision-and-recall-3da06bea9f6c
If you notice the way the recall is phrased within the terrorist example, it makes sense since what you want to identify is who is terrorist or not. It does not sound as informative if you just intend to catch a single terrorist (another example is if you use your biometric system to grant access to a single user). The point is that it is not wrong, but not as informative as the traditional measures used for the problem of identification.
When you use it to evaluate the way an ML clusters vectors into categories for faster lookups, then precision and recall become more interesting since it is measuring the impact on groupings as they are retrieved to see if the signatures are correctly within the groups.
Notice the difference between the two examples, in one I am more concerned with the individual and on the other I am more concerned with groups.
Arturo Geigel
Thanks very much for your help. Unfortunately, the link attached in your answer cannot be opened.
Asmaa Diab ,
I can open it, but it may be my browser session. If you want to Google
"towardsdatascience.com Beyond Accuracy: Precision and Recall" and you will get the link to it
Regards
- Badges
- Science topic
- Similar topics
- Computer Science
- Computing