Preliminary Exploration of Unknown Object Detection in Terahertz Radar Images through Incremental Recognition Method

Hidden object detection in terahertz radar images is crucial for public security and anti-terrorism efforts. Detecting unknown objects in real-world scenarios remains a challenging issue. To address this, we propose a novel study using incremental recognition methods. This approach trains with only a few samples from new object classes without needing to retrain old classes, reducing training time and cost while preserving recognition of previously learned objects. Within our research, we employ the Faster R-CNN architecture as the fundamental object detection model, subjecting it to training on both the base and new classes. To enhance training for new classes, we employ a new-class aware RPN to automatically label new class objects and use a fine-tuning strategy. Our experiments on the Active Terahertz Imaging Dataset for incremental recognition yielded good results. Our proposed method achieves a mean Average Precision (mAP) of 41.4% for recognizing all classes. Additionally, the recognition performance for new classes improved by 6.7% in Average Precision (AP) compared to simply incorporating new class training. This demonstrates the feasibility of incremental recognition for unknown object detection and validates the effectiveness of our proposed method.