Segmentação semantica com mixtura global de priors Gaussianas

Author(s)

Rodrigo Fumihiro de Azevêdo Kanehisa

• Document type: Master's Dissertation
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Instituto de Computação
• Defense date: 2022-04-27

Examining board members

Adín Ramírez Rivera; Hélio Pedrini; Roberto de Alencar Lotufo

• Advisor: Adín Ramírez Rivera

Abstract

Semantic segmentation is one of the open-ended problems in computer vision. This problem consists in classifying each pixel present in an image, delimiting an object. Convolutional neural networks are efficient in detecting and identifying, and segmenting objects in images. However, during the segmentation process, some problems arise, such as loss of spatial information and low resolution of feature maps. These problems lead to rough or noisy segmentations due to the uncertainties in the data. One way to counter these problems is to use statistical methods together with neural networks. This work proposes neural networks combined with a probabilistic module as representations for semantic segmentation focusing on medical images. Probabilistic models present a solution to map the contents of the image into more representative distributions on the latent space of the data, allowing the network to handle uncertainty and model more complex information. We evaluated our proposals within existing databases for image segmentation and compare them against existing methods on standard benchmarks for semantic segmentation. A more complex U-Net model will able more expressive, allowing the network to better reconstruct the information lost during the downsampling and stride operations (AU)

• FAPESP's process: 19/08589-0 - Semantic Segmentation based on Variational Methods
• Grantee: Rodrigo Fumihiro de Azevedo Kanehisa
• Support Opportunities: Scholarships in Brazil - Master