Date of Graduation


Document Type


Degree Type



Statler College of Engineering and Mineral Resources


Lane Department of Computer Science and Electrical Engineering

Committee Chair

Thirimachos Bourlai

Committee Co-Chair

Natalia Schmid

Committee Member

Natalia Schmid

Committee Member

Yuxin Liu

Committee Member

Frances Vanscoy


Deep learning models with convolutional neural networks are being used to solve some of the most difficult problems in computing today. Complicating factors to the use and development of deep learning models include lack of availability of large volumes of data, lack of problem specific samples, and the lack variations in the specific samples available. The costs to collect this data and to compute the models for the task of detection remains a inhibitory condition for all but the most well funded organizations. This thesis seeks to approach deep learning from a cost reduction and hybrid perspective — incorporating techniques of transfer learning, training augmentation, synthetic data generation, morphological computations, as well as statistical and thresholding model fusion — in the task of detection in two domains: visible spectrum detection of target spacecraft, and radio spectrum detection of radio frequency interference in 2D astronomical time-frequency data. The effects of training augmentation on object detection performance is studied in the visible spectrum, as well as the effect of image degradation on detection performance. Supplementing training on degraded images significantly improves the detection results, and in scenarios with low factors of degradation, the baseline results are exceeded. Morphological operations on degraded data shows promise in reducing computational requirements in some detection tasks. The proposed Mask R-CNN model is able to detect and localize properly on spacecraft images degraded by high levels of pixel loss. Deep learning models such as U-Net have been leveraged for the task of radio frequency interference labeling (flagging). Model variations on U-Net architecture design such as layer size and composition are continuing to be explored, however, the examination of deep learning models combined with statistical tests and thresholding techniques for radio frequency interference mitigation is in its infancy. For the radio spectrum domain, the use of the U-Net model combined with various statistical tests and the SumThreshold technique in an output fusion model is tested against a baseline of SumThreshold alone, for the detection of radio frequency interference. This thesis also contributes an improved dataset for spacecraft detection, and a simple technique for the generation of synthetic channelized voltage data for simulating radio astronomy spectra recordings in a 2D time-frequency plot.