Banca de QUALIFICAÇÃO: ANA PAULA PAGOTTI

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : ANA PAULA PAGOTTI
DATE: 01/03/2024
TIME: 14:00
LOCAL: Sala 406 do Bloco B do Campus de Santo André da Universidade Federal do ABC
TITLE:

Despeckling and Generating SAR Images throught Conditional GANs.

PAGES: 106
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Telecomunicações
SUMMARY:

The objective of this work is to generate Sentinel-1 SM ascending orbit SAR images, as this mode of acquisition of Sentinel-1 SAR sensor is more rare and are not avaiable to many geographic regions as IW ones. In the aim of augmenting this SM images dataset, after have been done an extensive research of which was a good technique to be applied, it was decided to use pix2pix network that is a very usable GAN. To do that, it is also created some new SAR datasets that includes pair of SM and IW images of the same imaged regions in both ascending and descending orbits respectively. These datasets comprehends regions with Rural, Urban and Coastal characteristics and both VH and VV polarizations. Althought, to the first tests presented here it is opted just to work only with Rural VH datasets. Also, as original pix2pix works with very small images, it was needed to modify its architecture to support bigger images and it is presented wich was the best adjustments done to improve its performance. In parallel it is also investigated the speckle removing problem to SAR images and the original images are threatened with classical despeckle filters as Lee, Lee-Sigma and Refined-Lee. Deep learning known methods are also applied to remove speckle as a previous trained network despeckNet and an the same one trained with the created dataset. Following, it is also proposed to use and train the network pix2pix as a despeckle filter to original images. Then, finally it is showed the tests of all these despeckle techniques applied to the generated images and it is also proposed a new approach where it is trained the pix2pix network to both generate and remove speckle at the same time, reducing the two steps of generating and filtering to only one. The results are all presented visually in images and both using quality parameters image metrics. The conclusion is that considering the computational efforts applied to develop all this network trains the better results are presente to the last proposed tecnhique of both generating and removng speckle at the same time using pix2pix, showing that this techinique should be more exploite.

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) Externo ao Programa - 2078059 - LUIZ ANTONIO CELIBERTO JUNIOR
Membro Suplente - Examinador(a) Interno ao Programa - 2356637 - KENJI NOSE FILHO