Melhoria de Qualidade de Imagens usando CNNs

Alan Peterson Ignácio Ralha Gonçalves, Gustavo de Oliveira Frade Duarte, Rafael Tadeu Cardoso dos Santos, Pedro de Carvalho Cayres Pinto, Gustavo Martins da Silva Nunes, Fernanda Duarte Vilela R. de Oliveira, Jose Gabriel Gomes

DOI: 10.14209/sbrt.2023.1570918201

Evento: XLI Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2023)

Keywords: Images Denoising CNN Autoencoder

Abstract

In this work, we analyze convolutional neural network (CNN) performances in image denoising. We trained autoencoder, residual autoencoder, and U-Net models to optimize 3-SSIM and SSIM quality metrics. SSIM improvements between 5 and 9 percentage points in comparison to the gaussian filter indicate that results obtained from the CNNs have higher similarity to the pristine images. We observe that deeper CNNs, with more parameters, tend to generate better quality images.

Download