Banca de DEFESA: Sana Alamgeer

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : Sana Alamgeer
DATE: 09/09/2022
TIME: 13:45
LOCAL: Sala de Seminários (Espaço Sérgio Barroso)
TITLE:

Deep Learning Based Objective Quality Assessment of Multidimensional Visual Content

KEY WORDS:

Visual Quality Assessment, 4D Light Fields, Visual Attention, Deep Learning, Bottleneck Features

PAGES: 153
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Eletrônica Industrial, Sistemas e Controles Eletrônicos
SPECIALTY: Automação Eletrônica de Processos Elétricos e Industriais
SUMMARY:

In the last decade, there has been a tremendous increase in the popularity of multimedia applications, hence increasing multimedia content. When these contents are generated, transmitted, reconstructed and shared, their original pixel values are transformed. In this scenario, it becomes more crucial and demanding to assess visual quality of the affected visual content so that the requirements of end-users are satisfied. In this work, we investigate effective spatial, temporal, and angular features by developing no-reference algorithms that assess the visual quality of distorted multi-dimensional visual content. We use machine learning and deep learning algorithms to obtain prediction accuracy. For two-dimensional (2D) image quality assessment, we use multiscale local binary patterns and saliency information, and train / test these features using Random Forest Regressor. For 2D video quality assessment, we introduce a novel concept of spatial and temporal saliency and custom objective quality scores. We use a Convolutional Neural Network (CNN) based light-weight model for training and testing on selected patches of video frames. For objective quality assessment of four-dimensional (4D) light field images (LFI), we propose seven LFI quality assessment (LF-IQA) methods in total. Considering that LFI is composed of dense multi-views, Inspired by Human Visual System (HVS), we propose our first LF-IQA method that is based on a two-streams CNN architecture. The second and third LF-IQA methods are also based on a two-stream architecture, which incorporates CNN, Long Short-Term Memory (LSTM), and diverse bottleneck features. The fourth LF-IQA is based on CNN and Atrous Convolution layers (ACL), while the fifth method uses CNN, ACL, and LSTM layers. The sixth LF-IQA method is also based on a two-stream architecture, in which, horizontal and vertical EPIs are processed in the frequency domain. Last, but not least, the seventh LF-IQA method is based on a Graph Convolutional Neural Network. For all of the methods mentioned above, we performed intensive experiments, and the results show that these methods outperformed state-of-the-art methods on popular quality datasets.

BANKING MEMBERS:
Externo à Instituição - CARLOS ALEXANDRE DE BARROS MELLO - UFPE
Externo à Instituição - CARLA LIBERAL PAGLIARI - IME
Interno - 3374036 - JOAO LUIZ AZEVEDO DE CARVALHO
Externo ao Programa - 1220587 - LI WEIGANG
Presidente - 1609346 - MYLENE CHRISTINE QUEIROZ DE FARIAS