Full text
| |
| Author(s): |
Augusto Cesar Monteiro Silva
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | São Paulo. |
| Institution: | Universidade de São Paulo (USP). Instituto de Matemática e Estatística (IME/SBI) |
| Defense date: | 2020-12-09 |
| Examining board members: |
Nina Sumiko Tomita Hirata;
Hae Yong Kim;
Letícia Rittner
|
| Advisor: | Nina Sumiko Tomita Hirata |
| Abstract | |
Manually designing an image operator that performs a specific transformation of images is a hard and time consuming task. The problem of automatically learning image operators has been researched throughout the years. Methods that tackle this problem can be roughly divided into three types: the traditional pixelwise or sliding-window approaches, the patch-to-patch approaches enabled by recent end-to-end deep learning models, and the structurally oriented approaches based on generative techniques. Each approach has its own advantages and drawbacks. The goal of this dissertation is to study the similarities and differences among these approaches, both conceptually and experimentally. In particular, we are interested in understanding how well structural information of the images such as connected thin lines are preserved. The first contribution of this work is an end-to-end method that joins the advantages of pixelwise and patch-to-patch approaches, which we call SConvNet. A second contribution is a study that shows that the skeletal similarity based metric is well suited for evaluating handwritten document binarization algorithms in a complementary way to traditional pixelwise metrics. At last, we present an experimental comparison among representative methods of the outlined three types of approaches, with respect to traditional pixelwise as well as the skeletal similarity metrics, on two image processing tasks (retinal blood vessel segmentation and handwritten document binarization). Better pixelwise metrics were achieved by patch-to-patch methods while better structural metrics were achieved by structural approaches. This is consistent with visual inspection, which shows that structural approaches better preserve the overall structure while patch-to-patch methods generate more precise contours. (AU) | |
| FAPESP's process: | 18/00477-5 - Learning image-to-image transformations |
| Grantee: | Augusto César Monteiro Silva |
| Support Opportunities: | Scholarships in Brazil - Master |