Date of Award

6-2-2025

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (Ph.D.)

First Advisor

Amir Shirkhodaie

Abstract

The collection and curation of high-quality data is crucial for developing robust and trustworthy deep learning (DL) models. Due to the costly and time-consuming nature of gathering well-annotated data, researchers have shifted to supplementing network training with synthetic data. However, when solely training AI classifiers with synthetic data, the network performance when exposed to data from real environments may not be satisfactory. To mitigate this shortcoming, a Gen-AI model was developed to systematically generate annotated, context-controlled synthetic imagery datasets using real SAR imagery. Through this process, SAR images of targets of interests are initially semantically segmented and introduced onto the designated regions of real SAR scene data. Additional blending of noise, environment objects, and obstruction objects are introduced into the scene to improve generated SAR image representation with dependable automated annotation. The proposed process allows accommodation of target objects into the scene in various orders of arrangement including random placement that results in one-of-a-kind contextual images with varying degrees of complication at each iteration that challenge the training of object detection models. This feature automatically facilitates production of large-scale generative SAR imagery with high annotation reliability. To test the fidelity of the generated data, a transfer learned Yolov8-OBB model was trained and tested on the scenes. The results show a 12.5% increase in mAP50-95 score, showing the network can classify objects at a higher confidence threshold than trained solely on the real data.

Download

Share

COinS