Date of Award
Electrical and Computer Engineering
Adaptive Support Weights, Background Suppression, Bilateral Filter, Gaussian Mixture Model, Image Segmentation, Video Segmentation
CC BY-NC-ND 4.0
Segmentation in images and videos has continuously played an important role in image processing, pattern recognition and machine vision. Despite having been studied for over three decades, the problem of segmentation remains challenging yet appealing due to its ill-posed nature. Maintaining spatial coherence, particularly at object boundaries, remains difficult for image segmentation. Extending to videos, maintaining spatial and temporal coherence, even partially, proves computationally burdensome for recent methods. Finally, connecting these two, foreground segmentation, also known as background suppression, suffers from noisy or dynamic backgrounds, slow foregrounds and illumination variations, to name a few. This dissertation focuses more on probabilistic model based segmentation, primarily due to its applicability in images as well as videos, its past success and mainly because it can be enhanced by incorporating spatial and temporal cues. The first part of the dissertation focuses on enhancing conventional GMM for image segmentation using Bilateral filter due to its power of spatial smoothing while preserving object boundaries. Quantitative and qualitative evaluations are done to show the improvements over a number of recent approaches. The later part of the dissertation concentrates on enhancing GMM towards foreground segmentation as a connection between image and video segmentation. First, we propose an efficient way to include multiresolution features in GMM. This novel procedure implicitly incorporates spatial information to improve foreground segmentation by suppressing noisy backgrounds. The procedure is shown with Wavelets, and gradually extended to propose a generic framework to include other multiresolution decompositions. Second, we propose a more accurate foreground segmentation method by enhancing GMM with the use of Adaptive Support Weights and Histogram of Gradients. Extensive analyses, quantitative and qualitative experiments are presented to demonstrate their performances as comparable to other state-of-the-art methods. The final part of the dissertation proposes the novel application of GMM towards spatio-temporal video segmentation connecting spatial segmentation for images and temporal segmentation to extract foreground. The proposed approach has a simple architecture and requires a low amount of memory for processing. The analysis section demonstrates the architectural efficiency over other methods while quantitative and qualitative experiments are carried out to show the competitive performance of the proposed method.
Mukherjee, Dibyendu, "Enhancing the Potential of the Conventional Gaussian Mixture Model for Segmentation: from Images to Videos" (2014). Electronic Theses and Dissertations. 5260.