Advanced search
Start date

Structural information in image-to-image transformation learning processes

Grant number: 19/07361-5
Support type:Scholarships abroad - Research Internship - Master's degree
Effective date (Start): August 01, 2019
Effective date (End): January 31, 2020
Field of knowledge:Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
Principal researcher:Nina Sumiko Tomita Hirata
Grantee:Augusto César Monteiro Silva
Supervisor abroad: Xiaoyi Jiang
Home Institution: Instituto de Matemática e Estatística (IME). Universidade de São Paulo (USP). São Paulo , SP, Brazil
Research place: University of Munster, Germany  
Associated to the scholarship:18/00477-5 - Learning image-to-image transformations, BP.MS


Image processing is an important step in many image analysis tasks. With deep neural networks, and particularly with the fully convolutional neural networks, one can learn image-to-image transformations from pairs of input-output training images. Although treating individual pixels as independent samples has shown impressive results, there is an understanding that models that take into consideration the spatial relationships between pixels are able to better capture structural information of image content. For instance, semantic segmentation networks in conjunction with conditional random fields(CRF) have shown state-of-the-art performance in image segmentation tasks. The aim of this project is to study such models, their optimization within the network optimization process, and develop improvements and new models to be integrated into general image-to-image transformation learning processes.