Publications
Detailed Information
Hybrid Approaches for MRF Optimization: Combination of Stochastic and Deterministic Methods
DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | 이경무 | - |
dc.contributor.author | 김원식 | - |
dc.date.accessioned | 2017-07-13T07:03:49Z | - |
dc.date.available | 2017-07-13T07:03:49Z | - |
dc.date.issued | 2014-02 | - |
dc.identifier.other | 000000018450 | - |
dc.identifier.uri | https://hdl.handle.net/10371/118996 | - |
dc.description | 학위논문 (박사)-- 서울대학교 대학원 : 전기·컴퓨터공학부, 2014. 2. 이경무. | - |
dc.description.abstract | Markov Random Field (MRF) models are of fundamental importance in computer vision. Many vision problems have been successfully formulated in MRF optimization. They include stereo matching, segmentation, denoising, and inpainting, to mention just a few. To solve them effectively, numerous algorithms have been developed. Although many of them produce good results for relatively easy problems, they are still unsatisfactory when it comes to more difficult MRF problems such as non-submodular energy functions, strongly coupled MRFs, and high-order clique potentials.
In this dissertation, several optimization methods are proposed. The main idea of proposed methods is to combine stochastic and deterministic optimization methods. Stochastic methods encourage more exploration in the solution space. On the other hand, deterministic methods enable more efficient exploitation. By combining those two approaches, it is able to obtain better solution. To this end, two stochastic methodologies are exploited for the framework of combination: Markov chain Monte Carlo (MCMC) and stochastic approximation. First methodology is the MCMC. Based on MCMC framework, population based MCMC (Pop-MCMC), MCMC with General Deterministic algorithms (MCMC-GD), and fusion move driven MCMC (MCMC-F) are proposed. Although MCMC provides an elegant framework of which global convergence is provable, it has the slow convergence rate. To overcome, population-based framework and combination with deterministic methods are used. It thereby enables global moves by exchanging information between samples, which in turn, leads to faster mixing rate. In the view of optimization, it means that we can reach a lower energy state rapidly. Second methodology is the stochastic approximation. In stochastic approximation, the objective function for optimization is approximated in stochastic way. To apply this approach to MRF optimization, graph approximation scheme is proposed for the approximation of the energy function. By using this scheme, it alleviates the problem of non-submodularity and partial labeling. This stochastic approach framework is combined with graph cuts which is very efficient algorithm for easy MRF optimizations. By this combination, fusion with graph approximation-based proposals (GA-fusion) is developed. Extensive experiments support that the proposed algorithms are effective across different classes of energy functions. The proposed algorithms are applied in many different computer vision applications including stereo matching, photo montage, inpaining, image deconvolution, and texture restoration. Those algorithms are further analyzed on synthetic MRF problems while varying the difficulties of the problems as well as the parameters for each algorithm. | - |
dc.description.tableofcontents | 1 Introduction 1
1.1 Markov random eld . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1.1 MRF and Gibbs distribution . . . . . . . . . . . . . . . . . . 1 1.1.2 MAP estimation and energy minimization . . . . . . . . . . . 2 1.1.3 MRF formulation for computer vision problems . . . . . . . . 3 1.2 Optimizing energy function . . . . . . . . . . . . . . . . . . . . . . . 4 1.2.1 Markov chain Monte Carlo . . . . . . . . . . . . . . . . . . . 7 1.2.2 Stochastic approximation . . . . . . . . . . . . . . . . . . . . 8 1.3 combination of stochastic and deterministic methods . . . . . . . . . 9 1.4 Outline of dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2 Population-based MCMC 13 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2.1 Swendsen-Wang Cuts . . . . . . . . . . . . . . . . . . . . . . 16 2.2.2 Population-based MCMC . . . . . . . . . . . . . . . . . . . . 19 2.3 Proposed Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.1 Segment-based stereo matching . . . . . . . . . . . . . . . . . 31 2.4.2 Parameter analysis . . . . . . . . . . . . . . . . . . . . . . . . 41 2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3 MCMC Combined with General Deterministic Methods 47 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.2 Related works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.3 Proposed algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.3.1 Population-based sampling framework for MCMC-GD . . . . 53 3.3.2 Kernel design . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 3.4.1 Analysis on synthetic MRF problems . . . . . . . . . . . . . . 60 3.4.2 Results on real problems . . . . . . . . . . . . . . . . . . . . . 75 3.4.3 Alternative approach: parallel anchor generation . . . . . . . 78 3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 4 Fusion Move Driven MCMC 89 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 4.2 Proposed algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 4.2.1 Sampling-based optimization . . . . . . . . . . . . . . . . . . 91 4.2.2 MCMC combined with fusion move . . . . . . . . . . . . . . . 92 4.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 5 Fusion with Graph Approximation 101 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 5.2 Related works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 5.2.1 Graph cuts-based move-making algorithm . . . . . . . . . . . 104 5.2.2 Proposals for fusion approach . . . . . . . . . . . . . . . . . . 106 5.3 Proposed algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 5.3.1 Stochastic approximation . . . . . . . . . . . . . . . . . . . . 107 5.3.2 Graph approximation . . . . . . . . . . . . . . . . . . . . . . 108 5.3.3 Overall algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 110 5.3.4 Characteristics of approximated function . . . . . . . . . . . 110 5.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 5.4.1 Image deconvolution . . . . . . . . . . . . . . . . . . . . . . . 113 5.4.2 Binary texture restoration . . . . . . . . . . . . . . . . . . . . 115 5.4.3 Analysis on synthetic problems . . . . . . . . . . . . . . . . . 118 5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 6 Conclusion 127 6.1 Summary and contribution of the dissertation . . . . . . . . . . . . . 127 6.2 Future works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 6.2.1 MCMC without detailed balance . . . . . . . . . . . . . . . . 128 6.2.2 Stochastic approximation for higher-order MRF model . . . . 130 Bibliography 131 국문초록 141 | - |
dc.format | application/pdf | - |
dc.format.extent | 12248005 bytes | - |
dc.format.medium | application/pdf | - |
dc.language.iso | en | - |
dc.publisher | 서울대학교 대학원 | - |
dc.subject | Markov random fields | - |
dc.subject | Combinatorial optimization | - |
dc.subject | Markov chain Monte Carlo | - |
dc.subject | Population based algorithm | - |
dc.subject | Stochastic approximation | - |
dc.subject | Non-submodular energy model | - |
dc.subject | Higher order energy model | - |
dc.subject.ddc | 621 | - |
dc.title | Hybrid Approaches for MRF Optimization: Combination of Stochastic and Deterministic Methods | - |
dc.type | Thesis | - |
dc.description.degree | Doctor | - |
dc.citation.pages | xvi, 142 | - |
dc.contributor.affiliation | 공과대학 전기·컴퓨터공학부 | - |
dc.date.awarded | 2014-02 | - |
- Appears in Collections:
- Files in This Item:
Item View & Download Count
Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.