PPGINF PÓS-GRADUAÇÃO EM ENGENHARIA DA INFORMAÇÃO FUNDAÇÃO UNIVERSIDADE FEDERAL DO ABC Teléfono/Ramal: No informado http://propg.ufabc.edu.br/ppginfo

Banca de DEFESA: ANA PAULA PAGOTTI

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : ANA PAULA PAGOTTI
DATE: 02/10/2024
TIME: 14:00
LOCAL: https://conferenciaweb.rnp.br/webconf/murilo-13
TITLE:

Speckle Filtering and Fake SAR Images Generation for Dataset Augmentation


PAGES: 196
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Telecomunicações
SUMMARY:

Radars are sensors that emit electromagnetic waves and capture their echo when they detect an obstacle. Among the various types of radar, this thesis will focus on Synthetic Aperture Radars (SAR). In general, these radars are installed on moving platforms, such as airplanes or satellites, and their images are not affected by weather conditions, such as the presence of clouds or lack of light. The SAR Sentinel-1 satellites, launched by the European Space Agency, provide images that are freely available; however, their ascending orbit SM acquisition mode, which has higher resolution, is not available for most regions in Brazil, differently from the descending orbit IW acquisition mode. Considering that current machine learning techniques for classification, target detection, and speckle noise removal require larger datasets to be properly trained than those available, the main objective of this thesis is to generate fake SAR images with ascending orbit characteristics from images acquired in descending orbit and to investigate filtering techniques that can be combined with the generation process. For this purpose, the Pix2pix generative adversarial neural network was selected and its architecture was modified to work with SAR images. In parallel, a study of speckle noise behavior in the generated fake images is presented, aiming to augment the filtering training dataset. The Pix2pix network is then applied as a speckle filter using both a training dataset with original Sentinel-1 SAR images, as well as this same dataset augmented by images previously generated by Pix2pix for training. The results of this proposed filter is compared to specific filtering techniques for SAR, such as the classical filters Lee, Lee-Sigma, and RefinedLee, as well as methods using other neural networks such as SAR-CNN and SAR2SAR. Finally, the speckle filters are also applied to the generated images and a new approach is also proposed, where the Pix2pix network is trained to generate and remove speckle simultaneously, reducing the two generation and filtering steps into a single one, optimizing computational resources. The results are presented visually by images that were evaluated by professionals in the area through a form using a scale adapted from NIIRS and by numerical image quality parameters. The conclusion is that although Pix2pix is a network with generic applicability, it is promising for the generation and increase of the set of SAR images and for speckle removal. It is also important to highlight that the proposed simultaneous SAR fake image generation and filtering brings good results, reducing computational costs and processing steps.


COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1761105 - MURILO BELLEZONI LOIOLA
Membro Titular - Examinador(a) Interno ao Programa - 2334927 - ANDRE KAZUO TAKAHATA
Membro Titular - Examinador(a) Interno ao Programa - 2356637 - KENJI NOSE FILHO
Membro Titular - Examinador(a) Externo ao Programa - 2078059 - LUIZ ANTONIO CELIBERTO JUNIOR
Membro Titular - Examinador(a) Externo à Instituição - RENATO MACHADO - ITA
Membro Suplente - Examinador(a) Interno ao Programa - 1761107 - RICARDO SUYAMA
Membro Suplente - Examinador(a) Externo à Instituição - RAFAEL KRUMMENAUER - UEM
Membro Suplente - Examinador(a) Externo à Instituição - DIMAS IRION ALVES - ITA
Notícia cadastrada em: 20/09/2024 20:19
SIGAA | UFABC - Núcleo de Tecnologia da Informação - ||||| | Copyright © 2006-2024 - UFRN - sigaa-1.ufabc.int.br.sigaa-1-prod