Two studies on Convolutional Neural Networks sensibility to resolution

• Ano de defesa: 2021
• Autor(a) principal: Abello, Antonio Augusto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Deep learning; Face recognition; Reconhecimento facial; Super resolution; Super-resolução
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-16122021-182010/

Convolutional Neural Networks (CNNs) recently became the state-of-the-art for various Computer Vision tasks. However, for reasons not completely understood, they are very sensitive to low resolution images. This can be troublesome as real life applications such as automated driving or surveillance can not use high resolution sensors. In this work we perform two studies on this subject matter: on the first we empirically study the effect of resolution loss and image restoration algorithms on a Face Recognition model. On the second, we study the high frequency bias hypothesis, one of the current possible explanations for CNNs sensitivity. We are able to develop new techniques for image restoration that better deal with the low resolution recognition problem and advance the understanding of the high frequency bias in CNNs.