Přístupnostní navigace
E-application
Search Search Close
Master's Thesis
Author of thesis: Ing. Ján Kačur
Acad. year: 2022/2023
Supervisor: Ing. Jakub Špaňhel, Ph.D.
Reviewer: Ing. Roman Juránek, Ph.D.
This thesis focuses on generating latent fingerprints using Generative adversarial networks. The main objective is to generate multiple latent fingerprints from the clean fingerprint, with the same identity. The identity and the style should also be controllable separately. The chosen approach is based on AugNet model. Designed algorithm generates latent fingerprints from clean binarized fingerprint, and a random vector encoding distortions, i.e style. In the generator, AdaIN blocks are used to incorporate distortions into the input fingerprint. Various training algorithms are tested, with WGAN-GP performing the best. Individual models are compared using a combination of FID, and Rank-1 accuracy on matching generated images to original input binarized fingerprints. Best performing models are selected as a Pareto optimal combinations of these 2 metrics.
fingerprint generation, latent fingerprint, GAN, conditional GAN, AugNet, MOLF, NIST SD302, WGAN-GP
Date of defence
16.06.2023
Result of the defence
Defended (thesis was successfully defended)
Grading
B
Process of defence
Student nejprve prezentoval výsledky, kterých dosáhl v rámci své práce. Komise se poté seznámila s hodnocením vedoucího a posudkem oponenta práce. Student následně odpověděl na otázky oponenta a na další otázky přítomných. Komise se na základě posudku oponenta, hodnocení vedoucího, přednesené prezentace a odpovědí studenta na položené otázky rozhodla práci hodnotit stupněm B.
Topics for thesis defence
Language of thesis
English
Faculty
Fakulta informačních technologií
Department
Department of Computer Graphics and Multimedia
Study programme
Information Technology and Artificial Intelligence (MITAI)
Specialization
Machine Learning (NMAL)
Composition of Committee
doc. Ing. Lukáš Burget, Ph.D. (předseda) doc. Ing. Martin Čadík, Ph.D. (člen) doc. Ing. Vladimír Janoušek, Ph.D. (člen) Ing. Michal Hradiš, Ph.D. (člen) Ing. Jaroslav Rozman, Ph.D. (člen) Ing. Tomáš Milet, Ph.D. (člen)
Supervisor’s reportIng. Jakub Špaňhel, Ph.D.
Práci hodnotím velmi pozitivně. Téma práce bylo opravdu náročné a dostupné metody velmi špatně zdokumentované. Student se práce chopil aktivně a po mnoha slepých uličkách doiteroval k funkčnímu řešení, které by mohlo sloužit jako základ generátoru syntetických dat z oblasti latentních otisků prstů pro průmysl. Celkově hodnotím práci stupněm A.
Tématem práce bylo generování latentních otisků prstů se zachováním identity zdrojového otisku.
Práci s literaturou hodnotím pozitivně. Student prostudoval doporučenou literaturu a následně si dohledal další zdroje důležité k dokončení práce.
Student konzultoval po celou dobu aktivně a pravidelně. Na konzultace chodil připraven a měl vždy hotový nějaký pokrok v řešení.
Práce byla mírně pozdržena nutností nagenerování otisků prstů po dokončení všech tréninků a následného zpracování za pomoci průmyslového partnera pro finální evaluaci. Vzhledem k množství natrénovaných modelů a složitosti evaluace je toto zdržení akceptovatelné. Text práce byl průběžně konzultován.
Student se zúčastnil konference Excel@FIT a za svoji práci získal i několik ocenění.
Grade proposed by supervisor: A
Reviewer’s reportIng. Roman Juránek, Ph.D.
The student surveyed a large number of methods for generating fingerprints using GAN models and performed a large number of quality evaluation experiments in cooperation with a commercial company. During the solution, he had to deal with a number of problems such as unclear formulation of the methods and thus had to invent many things himself. I positively assess that he was able to successfully develop the method for generating prints that may actually be of practical use. But it is a pity that he did not conduct an experiment that would prove its practical applicability.
Despite all the criticisms of the thesis text and source codes, I think the student showed his understanding of the topic and ability to successfully implement and evaluate neural network-based system.
Evaluation level: obtížnější zadání
I consider the topic of the thesis quite difficult due to bad documentation of the existing methods in literature. The student had to to orient himself in a large amount of contemporary methods which can be sometimes confusing.
The student focuses the thesis solely on latent fingerprints but there is no mention of the latent in the assignment. As I understood, generating clean fingerprints is not at this point interesting from industrial point of view and the focus is on the latent images due to lack of large scale data. In my view this is not problematic but just worth to mention here.
Evaluation level: zadání splněno
Evaluation level: je v obvyklém rozmezí
The work is divided into a theoretical and an experimental part, as is usual for such works. In the theoretical part, the student describes GAN models without reference to fingerprints. The FID metric is described here very briefly without context, thus it rather belongs to the experimental part. The subsection Fingerprint basics, in my opinion, belongs to the beginning of the thesis where it would make more sense. It is not clear from the text whether the student understands the described methods at the level at which he describes them or if he just took the descriptions from the articles.
In the experimental part, the student first describes the implementation of the methods he used and the way they were trained. Separate chapter is left for the experiments and their results. This part is structured logically and is quite readable. However, in my opinion, it would be better to avoid lengthy descriptions of what didn't work (or just briefly mention it) and focus on the final solution.
The thesis in nicely typeset. Some figures, however, have reduced resolution and missing important details. I did not notice any serious language issues.
Student cite relevant scientific literature regarding generative networks.
The uploaded files contain source codes for training models and the trained models. The sources are not commented, so it is very difficult to understand them. It is not even clear whether the student implemented everything himself (probably yes), or which parts from public open source projects were used. Codes for evaluation of experiments (calculation of Rank1 accuracy and summary of results) are missing.
The result is a set of models for generating latent fingerprints and source codes for their training. The generated images are not perfect and contain artifacts, which is expected due to the way they were generated. What I was missing was an experiment where the generated data would be used to train a real system for latent fingerprint recognition, in order to show whether the new data helps to increase the accuracy and robustness of the system (which is presumably the main reason for generating the data in the first place).
Grade proposed by reviewer: B
Responsibility: Mgr. et Mgr. Hana Odstrčilová