| Summary: | This dissertation explores the application of deep learning algorithms in the domain of Synthetic Aperture Radar (SAR) for Automatic Target Recognition (ATR), a field of increasing importance for military and surveillance purposes. Focusing on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset, the study provides an exhaustive comparison between two state-of-the-art convolutional neural network architectures: the You Only Look Once (YOLO) versions 5 and 8. The dissertation is motivated by the quest for a lightweight yet high-performing model capable of real-time deployment in resource-limited environments, such as embedded systems or FPGAs, and aims to navigate through the challenges posed by the unique characteristics of SAR imagery, such as speckle noise and geometric distortions, to achieve accurate and efficient target recognition.
We examine the intricacies of SAR image feature extraction, the limitations posed by scarce datasets, and the intra- and inter-class variations that affect classification. We also investigate the YOLO architecture known for its compactness, minimal parameterization, and resistance to noise—traits desirable for real-time recognition. A series of experiments compare the YOLOv5 and YOLOv8 series across precision, recall, and mean average precision (mAP) metrics.
The result shows that the YOLOv8 series marks a significant improvement over YOLOv5, offering a balanced, accurate, and robust solution for SAR ATR tasks. It is well-positioned to meet the demands of accuracy and computational efficiency required for real-time and noise-resistant applications. Future work is encouraged to optimize the YOLO model's deployment in constrained environments, enhance noise resistance, and explore its transferability and adaptability for varied ATR tasks.
|
|---|