New marked point process models for microscopy images
In developing new materials, the characterization of microstructures is one of the key steps. To characterize the microstructure, many microscope modalities have been devised and improved over decades. With the increase in image resolution in the spatial and time domains, the amount of image data keeps increasing in the fields such as materials science and biomedical engineering. As a result, image processing plays a critical role in this era of science and technology. In materials image analysis, image segmentation and feature detection are considered very important. The first part of this research aims to resolve the segmentation problem caused by blurring artifacts in scanning electron microscopy (SEM) images. This blurring issue can lead to a bridged channel problem, which becomes an obstacle in analyzing the microstructures. To tackle the problem, we propose a joint deconvolution and segmentation (JDS) method. As a segmentation method, we use the expectation-maximization/maximization of the posterior marginals (EM/MPM) method, using the Markov random field (MRF) prior model. Experiments show the proposed method improves the segmentation result at object boundaries. The next phase of the image segmentation is detecting image features. In the second part of this research, we detect channel configurations in materials images. We propose a new approach of channel identification, based on the marked point process (MPP) framework, to effectively detect channels in materials images. To describe a higher level of structures in an image, the MPP framework is more effective than the MRF prior model. The reversible-jump Markov chain Monte Carlo (RJMCMC) algorithm embedded with simulated annealing is used as an optimization method, and a new switching kernel in an RJMCMC is used to reduce computational time. The channel configuration is useful in characterizing materials images. In addition, this information can be used to reduce the bridged channel problem more effectively. In materials image processing, one of the most important goals of feature detection is identifying the 3D structure of objects from 3D microscope datasets. The final part of this research is to perform fast and accurate estimation of 3D object configurations from a 3D dataset. We propose a fast 3D fitting method to improve the computational complexity over a full-search 3D MPP method. Experiments show that the fast 3D fitting method significantly decreases execution time compared to the full 3D MPP method.
Comer, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our