Banca de QUALIFICAÇÃO: LEONARDO ALVES FERREIRA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : LEONARDO ALVES FERREIRA
DATE: 28/06/2023
TIME: 16:00
LOCAL: https://conferenciaweb.rnp.br/webconf/andre-39
TITLE:

Deep Image Prior for Deblurring and Tomographic Image Reconstruction

PAGES: 50
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Telecomunicações
SUMMARY:

The successful use of Artificial Neural Networks (ANN) has been revolutionizing many areas of human life in the past years. However, obtaining good results using ANNs is usually hindered by the need for a great amount of high-quality data. Recently, a new technique named Deep Image Prior (DIP) was proposed for image processing applications. In this technique, a generative ANN is used without the need for a training dataset. The main idea is that the ANN structure provides prior information that favors the generation of natural images, benefiting the obtained results. Since its proposal, the use of DIP has been demonstrated in many applications. In some cases, its results surpassed traditional methods and were close to the performance of ANNs that were trained with big datasets. Based on this context, this thesis aimed to explore the use of DIP for deblurring and tomographic reconstructions. Specifically, we investigated the performance of the technique for experimental data presenting severe quality problems, namely digital photographs with a high degree of defocus blur and Computed Tomography images acquired with a very limited angle. We also developed an unprecedented application of DIP on Electrical Impedance Tomography. Furthermore, we proposed and implemented additions to the technique to improve its performance. Our results demonstrated that DIP can provide adequate results even for poor-quality input data. Also, we demonstrated that the method can be further improved with additional modifications. Some limitations were also identified, such as unsatisfactory results in the most severely impaired cases of data, and the long time to generate each image. Nonetheless, DIP demonstrated to be a promising method for the image processing area, presenting many possibilities for further investigation regarding its improvement and its understanding.

COMMITTEE MEMBERS:
Presidente - 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 - 2123676 - OLAVO LUPPI SILVA
Membro Suplente - Examinador(a) Interno ao Programa - 1761105 - MURILO BELLEZONI LOIOLA