Master Thesis Defense - April 17, 2019: Color Image Segmentation by Integrating Spatial Information using Semi-Bounded Finite Mixture Models

Jaspreet Singh Kalsi

Wednesday, April 17, 2019 at 10:00 a.m.
Room EV003.309

You are invited to attend the following M.A.Sc. (Information Systems Security) thesis examination.

Examining Committee

Dr. W. Lucia, Chair
Dr. N. Bouguila, Supervisor
Dr. J. Bentahar, CIISE Examiner
Dr. F. Shadmehri, External Examiner (MIAE)

Abstract

In computer vision, image segmentation plays foundational role. Innumerable techniques, such as active contour, graph-cut-based, model-based, machine learning, and clustering-based methods have been proposed for tackling the image segmentation problem. But, none of them is universally applicable. Thus, the hunt for optimized and robust models for image segmentation is still under-process and also an open question. The challenges faced in image segmentation are the integration of spatial information, finding the exact number of clusters (M), and to segment the image smoothly without any inaccuracy specially in the presence of noise, a complex background, low contrast and, inhomogeneous intensity. The use of finite mixture model (FMMs) for image segmentation is very popular approach in the field of computer vision. The application of image segmentation using FMM ranges from automatic number plate recognition, content-based image retrieval, texture recognition, facial recognition, satellite imagery etc. Image segmentation using FMM undergoes some problems. FMM-based image segmentation considers neither spatial correlation among the peer pixels nor the prior knowledge that the adjacent pixels are most likely belong to the same cluster. Also, color images are sensitive to illumination and noise. To overcome these limitations, we have used three different methods for integrating spatial information with FMM. First method uses the prior knowledge of M. In second method, we have used Markov Random Field (MRF). Lastly, in third, we have used weighted geometric and arithmetic mean template. We have implemented these methods with inverted Dirichlet mixture model (IDMM), generalized inverted Dirichlet mixture model (GIDMM) and inverted Beta Liouville mixture model (IBLMM). For experimentation, the Berkeley 500 (BSD500) and MIT's Computational Visual Cognition Laboratory (CVCL) datasets are employed. Furthermore, to compare the image segmentation results, the outputs of IDMM, GIDMM, and IBLMM are compared with each other, using segmentation performance evaluation metrics.