Abstract

This dissertation presents a study of various machine learning techniques for recognizing vehicular objects in infrared images. State of the art methods for computer vision have not been widely explored for this part of the electromagnetic spectrum (EM). Challenges that arise due to the dearth of infrared training images, terrain clutter, and thermal phenomenology have not been fully addressed. Infrared dataset collection and annotation is both difficult and expensive. What if there is a way we can generate infrared images and diminish the need for collecting data out in the field? Our first research study encompasses an encoder-decoder model that predicts how objects will look like in novel views (azimuth) and orientation. Following this, our second study is another attempt to leverage on few infrared images for few shot learning (FSL). We exploit a traditional convolution neural network by designing filters for its first layers using clustering and other statistical representations that best describe the training data. Our method enables us to lay down class boundaries on the manifold such that the class distinction between them is maximized. An unseen class with very less training images is subsequently learnt with high classification accuracy. It is well known that deep networks for object classification can be sensitive to the image background, and may not always learn to focus on the semantically important regions of the image. In our next study, we propose new methods to solve this problem for the infrared domain. First, we present a deep learning based method called 'Split-training' which focuses on learning object activations only using masked data and transferring the learnt representations to unmasked images. Secondly, we propose a minimum mean square error (MMSE) distance classifier that learns from synthetic data only and focuses on the object while making a prediction.

Notes

If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu

Graduation Date

2022

Semester

Fall

Advisor

Mahalanobis, Abhijit

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Degree Program

Computer Science

Format

application/pdf

Identifier

CFE0009330; DP0027053

URL

https://purls.library.ucf.edu/go/DP0027053

Language

English

Release Date

December 2022

Length of Campus-only Access

None

Access Status

Doctoral Dissertation (Open Access)

Download

Share

COinS