- Title Information
- Title
- New Algorithms for Appearance Modeling in Image Segmentation
- Type of Resource (primo)
- dissertations
- Name:
Personal
- Name Part
- Farias Sales Rocha Neto, Jeova
- Role
- Role Term:
Text
- creator
- Name:
Personal
- Name Part
- Felzenszwalb, Pedro
- Role
- Role Term:
Text
- Advisor
- Name:
Personal
- Name Part
- Harrison, Matthew
- Role
- Role Term:
Text
- Reader
- Name:
Personal
- Name Part
- Kimia, Benjamin
- Role
- Role Term:
Text
- Reader
- Name:
Corporate
- Name Part
- Brown University. Engineering: Electrical Sciences and Computer Engineering
- Role
- Role Term:
Text
- sponsor
- Origin Information
- Copyright Date
- 2021
- Physical Description
- Extent
- 15, 122 p.
- digitalOrigin
- born digital
- Note:
thesis
- Thesis (Ph. D.)--Brown University, 2021
- Genre (aat)
- theses
- Abstract
- Modeling pairwise pixel interactions and how image segments can be distinguished from one another are common starting points of many segmentation approaches. In this thesis, we propose new formulations and methods for classical image segmentation algorithms based on modeling the image's appearance.
We first tackle the Markov Random Field (MRF) formulation of image segmentation, which requires modeling the typical color values of each desired segment prior to the segmentation algorithm. Traditionally, MRF-based segmentation is approached via iterative model estimation-segmentation algorithms. Here, we propose a novel framework that is able to estimate these models directly without user intervention or iterations. That is accomplished by making use of higher-order image color statistics. For the case of binary segmentation, the proposed estimation algorithms apply simple algebraic manipulations on mathematical expressions of second-order color moments. We also propose a model estimator for the case of multiregion images.
We also made contributions to modeling pairwise pixel interactions for spectral image segmentation. Here, we proposed new graph formulations that allow for long range edges, connecting pixels potentially far from each other, but that still present high appearance similarity. Within this new perspective, we build two graphs that incorporate different valuable segmentation cues: one that promotes the local cohesion of image segments and the other that encodes the global appearance dependencies among pixels. Differently from traditional spectral segmentation methods, we show that clustering each proposed graph separately advances a relevant segmentation cue. Computationally, we also provide graph sparsification algorithms that tackle our methods' computational requirements.
Finally, we also introduce an algorithmic framework for adding segmentation cues and soft constraints to the traditional spectral segmentation formulation. Our method is a generalization of the the Normalized Cuts algorithm via the incorporation of a plug-and-play penalty functions that codify prior segmentation knowledge. This new cut criterion can be approximately optimized via the solution of a sparse generalized eigenvector problem. We show that our proposed framework can be used as an alternative method to incorporate user-provided segmentation hints in the form of seeds. Our preliminary results show the effectiveness of our methods in some synthetic and real segmentation scenarios.
- Subject (fast)
(authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/00967482")
- Topic
- Image analysis
- Subject
- Topic
- Image Segmentation
- Subject (fast)
(authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/01129072")
- Topic
- Spectral theory (Mathematics)
- Subject (fast)
(authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/01012085")
- Topic
- Mathematical models
- Subject
- Topic
- Graph Cuts
- Language
- Language Term (ISO639-2B)
- English
- Record Information
- Record Content Source (marcorg)
- RPB
- Record Creation Date
(encoding="iso8601")
- 20211004