Back to results list
Please use this identifier to cite or link to this item:
|Title:||The analysis of invariant repetitive patterns in images and videos||Authors:||Cai, Yunliang||Keywords:||Image processing -- Digital techniques.
Hong Kong Polytechnic University -- Dissertations
|Issue Date:||2013||Publisher:||The Hong Kong Polytechnic University||Abstract:||The efficient and robust extraction of repetitive patterns from images is a ground challenge problem in computer vision. The repetitive patterns in images are products of both repetitive structures as well as repetitive reflections or color patterns. The perception of repetitive patterns in an image is strongly linked to the visual interpretation and composition of objects or textures. In other words, patterns that exhibit near-stationary behavior provide rich information about objects, their shapes, and their textures in an image. However, methodologies for repetition detection are still limited and there are few methods for analyzing the repetition composition. For the repetitive pattern detection and compositional analysis, there are three fundamental problems which we need to solve: 1. How to identify and extract repetitive invariant patterns (i.e., textons, texels, voxels) from images and videos. 2. How to cluster and group the repetitive patterns according to their intrinsic geometry. 3. How to construct the composition model for the detected repetitive patterns, when the patterns can be organized in a global structure. To answer the first and the second questions, we propose a new algorithm for repetitive pattern detection and grouping. The algorithm follows the classical region growing image segmentation scheme. A modified mean-shift algorithm is utilized to group local image patches into clusters. A continuous joint alignment is exploited to match similar patches and refine the subspace grouping.
To answer the third question, we propose an algorithm for inferring the composition structure of the repetitive patterns. The inference algorithm is based on the data-driven structural completion fields and the compositional analysis of structure generators. The global structure is then recovered by identifying the types of generators involved and the corresponding generator composition. As an extension of our repetitive pattern analysis for images, additional investigations are included for the extraction of the repeated spatial-temporal patterns in videos. We propose a low-dimensional registration method that projects the optical flow patterns into two dimensional space then aligns them onto this space. The low-dimensional registration is formulated as a point matching problem in the space of symmetric positive definite matrixes with spatial constraints. This can be efficiently solved by a loopy belief propagation. The proposed pattern detection and higher-level grouping are tested on the PSU-NRT dataset, obtaining competitive results of repetition detection and structure inference. We show the application of our method for the inference of the geometry of objects and the estimation for the general layout of a crowded scene. The structure inference can help us analyze the geometric structure of textures, and it can be applied to texture retrieval, shape from structure, and scene recognition.
|Description:||xxvi, 146 leaves : col. ill. ; 30 cm.
PolyU Library Call No.: [THS] LG51 .H577P COMP 2013 Cai
|URI:||http://hdl.handle.net/10397/6221||Rights:||All rights reserved.|
|Appears in Collections:||Thesis|
Show full item record
Files in This Item:
|b26392343_link.htm||For PolyU Users||203 B||HTML||View/Open|
|b26392343_ir.pdf||For All Users (Non-printable)||7.16 MB||Adobe PDF||View/Open|
Citations as of Oct 15, 2018
Citations as of Oct 15, 2018
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.