Self-organizing approach to learn a level-set function for object segmentation in complex background environments

Date of Award

2015

Degree Name

Ph.D. in Electrical Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Advisor: Vijayan K. Asari

Abstract

Boundary extraction for object region segmentation is one of the most challenging tasks in image processing and computer vision areas. The complexity of large variations in the appearance of the object and the background in a typical image causes the performance degradation of existing segmentation algorithms. One of the goals of computer vision studies is to produce algorithms to segment object regions to produce accurate object boundaries that can be utilized in feature extraction and classification. This dissertation research considers the incorporation of prior knowledge of intensity/color of objects of interest within segmentation framework to enhance the performance of object region and boundary extraction of targets in unconstrained environments. The information about intensity/color of object of interest is taken from small patches as seeds that are fed to learn a neural network. The main challenge is accounting for the projection transformation between the limited amount of prior information and the appearance of the real object of interest in the testing data. We address this problem by the use of a Self-organizing Map (SOM) which is an unsupervised learning neural network. The segmentation process is achieved by the construction of a local fitted image level-set cost function, in which, the dynamic variable is a Best Matching Unit (BMU) coming from the SOM map.The proposed method is demonstrated on the PASCAL 2011 challenging dataset, in which, images contain objects with variations of illuminations, shadows, occlusions and clutter. In addition, our method is tested on different types of imagery including thermal, hyperspectral, and medical imagery. Metrics illustrate the effectiveness and accuracy of the proposed algorithm in improving the efficiency of boundary extraction and object region detection. In order to reduce computational time, a lattice Boltzmann Method (LBM) convergence criteria is used along with the proposed self-organized active contour model for producing faster and effective segmentation. The lattice Boltzmann method is utilized to evolve the level-set function rapidly and terminate the evolution of the curve at the most optimum region. Experiments performed on our test datasets show promising results in terms of time and quality of the segmentation when compared to other state-of-the-art learning-based active contour model approaches. Our method is more than 53% faster than other state-of-the-art methods. Research is in progress to employ Time Adaptive Self- Organizing Map (TASOM) for improved segmentation and utilize the parallelization property of the LBM to achieve real-time segmentation.

Keywords

Image segmentation, Self-organizing systems, Image processing, Computer Engineering, Electrical Engineering, active contour models, level set function, self organizing map, lattice Boltzmann method, object segmentation

Rights Statement

Copyright © 2015, author

Link to Full Text

Share

COinS