Computational Approaches towards Pattern Recognition and Matching for Scientific Images
Open Access
- Author:
- Zheng, Xinye
- Graduate Program:
- Information Sciences and Technology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- April 15, 2020
- Committee Members:
- James Z Wang, Dissertation Advisor/Co-Advisor
James Z Wang, Committee Chair/Co-Chair
Anna Cinzia Squicciarini, Committee Member
Saeed M Abdullah, Committee Member
Chris Eliot Forest, Outside Member
Jia Li, Committee Chair/Co-Chair
Mary Beth Rosson, Program Head/Chair
Jia Li, Dissertation Advisor/Co-Advisor - Keywords:
- scientific image processing
optimal transport
machine learning
scientific image processing
machine learning
optimal transport
statistical modeling - Abstract:
- Pattern recognition and pattern matching are two important tasks in scientific image analysis. Researchers use patterns on image data to observe the phenomena that cannot be captured by human eyes, to model the activities of certain objects, and to evaluate the mathematical and physical assumptions. Tasks of pattern recognition and pattern matching on scientific images are complicated because they require domain knowledge to interpret the patterns. Compared with standard photography, scientific images usually depict a fussy state which is not self-explainable in most cases. In addition, the increasing size of image data in this big data era compounds the workload for researchers in processing and analyzing these data. This thesis is motivated to construct faster and more accurate computational models for scientific images using machine learning and optimization techniques. Instead of seeking a unified processing tool for all scientific images, this thesis proposes computational approaches according to the actual requirements and challenges in two real-world problems. In the first part, this thesis employs pattern recognition to detect severe weather events early. While numerical weather prediction is widely used in the weather forecast, we still rely on people to interpret high-level visual clues from satellite and radar visualizations. A machine-learning-based approach is proposed to detect any "comma-shaped cloud", which is a specific cloud distribution pattern strongly associated with storm formation. To train the model, we extracted shape and motion-sensitive features from delicately selected regions. The validated utility and accuracy of our method suggest a high potential for assisting meteorologists in weather forecasting. In the second part, this thesis proposes problem-specific pattern matching frameworks on scientific image processing applications under the settings of optimal transport (OT). OT is a widely used optimization tool that can solve the minimal transportation cost between two probability distributions. The classical OT formula, however, asserts strict mass transportation constraints, which is not feasible for matching the complicated patterns in scientific images. Additionally, because the standard OT setup optimizes minimal spatial distance, it may not work on the special data structures of scientific images. Scientific images can match according to other features, including shape, scale, color, and contours. To overcome the limitations of standard OT, we use the geometry of data structure to regulate OT and design computational models according to the requirement of applications. For example, for the task to match cells on microscopy, a bipartite graph is proposed to regularize OT and combine cell shape and location information in the optimization function. For another task to match two color distributions in image editing, Gaussian mixture models (GMM) are used to capture the color semantics better. Instead of forcing the mapping between two discrete color histograms, this thesis proposes a parametric framework to remove the artifact. This thesis concludes by discussing some future directions in developing more intelligent pattern recognition and matching systems for interpreting scientific images.