Analysis of the robustness and sensitivity of deep learning models for automatic target recognition.

Automatic Target Recognition (ATR) is an essential technology in various domains such as security, autonomous vehicles, and military applications. ATR systems are designed to detect, recognize, and classify objects of interest using sensor data, typically infrared imagery, to ensure reliable real-time operations under varying environmental conditions. However, the robustness and accuracy of ATR systems are often challenged by factors such as target occlusion, varying distances, and environmental perturbations. This thesis investigates the application of state-of-the-art deep learning techniques to enhance the performance and robustness of ATR systems. Specifically, we explore the use of Vision Transformers (ViT), Swin Transformer, and a Baseline VGG model to address the challenges of target recognition in complex scenarios. Additionally, two novel ViT variants are proposed, focusing on improving target localization through the use of larger image patches with varying degrees of overlap. The models are evaluated on infrared datasets under realistic conditions where targets may be poorly localized, occluded, or affected by noise. The performance of these models is systematically assessed by analyzing their resilience to input data fluctuations, including scaling, shifting, and other perturbations. The key research questions addressed in this thesis are: \\(1) Which deep learning classification model is most accurate for ATR tasks? \\(2) Which model demonstrates the highest robustness to variations and imperfections in the input data? Our findings indicate that while all models offer significant advancements over traditional ATR methods, the proposed ViT variants show exceptional promise in maintaining accuracy under challenging conditions. This research contributes to the development of more reliable and efficient ATR systems capable of robust performance in real-world applications.