Banca de DEFESA: VICTOR HUGO DE AGUIAR ARRUDA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : VICTOR HUGO DE AGUIAR ARRUDA
DATE: 25/04/2023
TIME: 09:30
LOCAL: Sessão por videoconferência. Google Meet/Sala: https://meet.google.com/hoy-hvic-mgo
TITLE:

PORE DETECTION IN FINGERPRINT TRACE IMAGES USING CONVOLUTIONAL NEURAL NETWORKS .

KEY WORDS:

Fingerprint traces. Convolutional neural network. Grid Search. True and false detection rate.

PAGES: 74
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

In this work, techniques that allow the detection of pores in fingerprint images are presented and evaluated. The first evaluation takes place in relation to the techniques that use filtering tools for the pore extraction and detection process. The second, in turn, concerns the use of convolutional neural networks (CNN), with a focus on machine learning. After comparing the existing techniques, the neural network topology proposed by (ALI; WANG; AHMAD, 2021) was implemented. The choice was motivated by the fact that the method in question guarantees the best results compared to the others, taking into account the metrics called true detection rate (RT) and false detection rate (RF). Such metrics are used to verify the success and error rate of the network when detecting a set of pores in the fingerprint image, respectively. As you can imagine, the network will be better the higher the hit rate and the lower the error rate. A contribution that is also significant in this work, and that deserves to be highlighted, is the use of a database, which, as far as is known in the researched literature, has never been used. This database is composed of real images obtained through forensic photography on images of fin- gerprint traces found at a crime scene. In addition, this master’s thesis innovated in the sense of using images with a resolution of 96 dpi, which is much lower than the resolution of 800 dpi established as a minimum for this pore detection process. To finalize the contributions, we must also talk about the implementation of Grid Search as a tool for choosing and optimizing neural network hyperparameters. The results obtained after the previously described implementations reveal that, initially, the rates of RTand RFwere not very satisfactory, reaching values of RF= 77,9% and RT= 56,5%, respectively. However, when the window analysis was performed, which corresponds to the restriction of the analysis on subimages of the main image, it was verified that, for a window of size 500 × 200, the algorithm presented satisfactory results, reaching a RTof 80,8% and an RFof 4,5%, respectively. In the end, the use of the method proposed in this work proved to be efficient in the process of detecting pores in images of traces of fingerprints found at a crime scene, a fact that contributes significantly in the field of Forensic Sciences, since with the marking from a minimum of 20 pores it is possible to individualize a person. 

COMMITTEE MEMBERS:
Externo à Instituição - FRANCISCO MADEIRO BERNARDINO JUNIOR - UPE
Externo ao Programa - 1055817 - JOAO MARCELO XAVIER NATARIO TEIXEIRA - UFPEPresidente - 1882484 - JULIANO BANDEIRA LIMA