Skip to main content
eScholarship
Open Access Publications from the University of California

UCLA

UCLA Electronic Theses and Dissertations bannerUCLA

Improving Image Feature Detection and Classification in Low-Label Regime with Deep and Classical Methods

Abstract

In today's world, data is fundamental to the growth of technology. However, it is very common that vast amounts of clean, well annotated data are unavailable and so there is a desire for methods that can function with small amounts of labeled data. This includes methods that utilize hand-crafted features and can be tuned manually for a dataset without having a variety of samples, as well as machine learning methods that can succeed in a low label rate regime by utilizing mutual information. This thesis discusses work involving hand-crafted features and variational methods for image detection, segmentation, and classification of shape-coded medical testing particles. Specifically, the Canny edge detector, Hough Transforms, and snake active contours are used for the object detection and segmentation problem, which offers a strong alternative to deep learning methods and has theoretical guarantees. Furthermore, this thesis explores neural networks, graph-based learning, and active learning methods that are robust to low label environments with applications in remote sensing and hyperspectral data. These methods offer a natural fusion of deep learning methods with transductive, variational label propagation. Altogether, these thesis offers a survey of image processing as a field and showcases a wide range of ideas that are all applicable for different problems.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View