Processing Image Data from Unstructured Environments
Kasapis, Spiridon
2023
Abstract
Advanced mobility research centers capture large amounts of data from ground vehicle systems during development and experimentation in both manned and autonomous operations. This exponential growth of digital image data has given rise to the need of understanding the content of image datasets by clustering and classifying them without the use of manual labor. Currently, there is a lack of tools which -through processing raw data- can provide a semantic understanding of an environment or dataset and can be used in place of a human to provide context to situations that threaten the uninterrupted operation of an autonomous vehicle. In search, exploration, and reconnaissance tasks performed with autonomous ground vehicles, an image classification capability is needed for specifically identifying targeted objects (relevant classes) and at the same time recognize when a candidate image does not belong to anyone of the relevant classes (irrelevant images). An open-set low-shot (OSLS) classifier was developed for addressing this need. During its training, it uses a modest number (less than 40) of labeled images for each relevant class, and unlabeled irrelevant images that are randomly selected at each epoch of the training process. The new OSLS classifier is capable of identifying images from the relevant classes, determining when a candidate image is irrelevant, and it can further recognize categories of irrelevant images that were not included in the training (unseen). The OSLS was integrated with an unsupervised learning feature extraction framework based on the instance discrimination method for creating an instance discrimination low shot (IDLS) module. The IDLS can identify targeted objects while at the same time recognize when candidate images do not belong to any one of the target classes, both in a very data-inexpensive way. The IDLS is dynamic, adapts to new environments during operation and is resilient to adversaries. The OSLS and IDLS algorithms were compared to a variety of alternative supervised methods showing comparable and often times better results in performing classification tasks, while requiring very few labeled images for training (i.e. less that 0.3% of labeled data compared to a supervised CNN for comparable levels of accuracy). This work also developed a soft-labeling capability for grouping collected images into categories using a new formulation that is based on an extended variance ratio criterion (E-VRC). The E-VRC comprises an unsupervised clustering capability since it does not require any initializations or prior knowledge about how many clusters will be encountered. As it is done with the previous two modules (OSLS and IDLS), the E-VRC too is being tested on several different datasets, demonstrating that it is useful not only for autonomous exploration and reconnaissance operations but also for the efficient content management and retrieval tasks. Additionally, the E-VRC algorithm developed by this research was compared to other available unsupervised clustering methods yielding superior results.Deep Blue DOI
Subjects
Unsupervised Training Computer Vision
Types
Thesis
Metadata
Show full item recordCollections
Remediation of Harmful Language
The University of Michigan Library aims to describe library materials in a way that respects the people and communities who create, use, and are represented in our collections. Report harmful or offensive language in catalog records, finding aids, or elsewhere in our collections anonymously through our metadata feedback form. More information at Remediation of Harmful Language.
Accessibility
If you are unable to use this file in its current format, please select the Contact Us link and we can modify it to make it more accessible to you.