This thesis investigates the application of advanced machine learning techniques for sub-terahertz imaging under limited data constraints. A detailed exploration of the impact of these constraints on object detection is presented. This study highlights the challenges faced by object detectors when trained with imbalanced or insufficient datasets, thus emphasizing the necessity for more advanced machine learning methods.

In response to the challenge posed by imbalanced data, this thesis evaluates the effectiveness of fine-tuning Few-Shot Object Detection frameworks for training the Faster-RCNN model, focusing on the scenario with limited image data of dangerous concealed items. To address the domain gap issue when adapting these frameworks to sub-terahertz images, this thesis proposes an innovative framework for generating high-quality pseudo-annotations from unlabeled datasets, thereby enhancing the object detector's performance.

To address the challenge of using insufficient datasets, this thesis introduces a novel framework that employs visual anomaly detection methods, which require minimal annotation for the training process. The proposed framework achieves a detection performance comparable to that of fully supervised detectors trained on a limited range of concealed objects, demonstrating the significant potential for (sub-)terahertz imaging applications.