Learning image transformations via classification of micro-regions

• Grant number: 18/11899-8
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: September 01, 2018
• End date: August 31, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Nina Sumiko Tomita Hirata
• Grantee: Pedro Henrique Barbosa de Almeida
• Host Institution: Instituto de Matemática e Estatística (IME). Universidade de São Paulo (USP). São Paulo , SP, Brazil

Abstract

Several types of image transformations can be performed by image operators modeled as functions that process each pixel individually. This characteristic allows one to insert the problem of determining these operators into the machine learning context as a problem of learning pixel classifiers. Among the image transformations, one that stands out is the segmentation, which produces a partition of the image points where some regions correspond to components of interest while others are those to be discarded in subsequent analysis steps. Due to the high computational cost of individual pixel processing, similar pixels could be grouped together into micro-regions to reduce the number of atomic components. Then, the problem of classifying pixels could be replaced with the problem of classifying these micro-regions. In segmentation problems, as far as the contours of these micro-regions present good adherence to the contours of interest, there is no loss of precision and there might be a significative gain regarding computational cost. The goal of this research project is to extend the set of existing image transform learning methods by developing new methods that learn transformations that act on micro-regions. The new methods will be integrated to TRIOSlib, a library maintained by the group, applied on different image segmentation tasks, and then compared with image transforms that act on pixels.