Passive terahertz image detection

• Main Author: Liu, Jikun
• Other Authors: Kai-Kuang Ma
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/167073
• Institution: Nanyang Technological University

Passive terahertz (THz) devices enable detection and imaging of concealed objects without the need for any radiation emitter. This project presents a specialized solution for the automated, precise, and real-time identification of hidden objects from passive terahertz images affected by random noise and significant striping, which hindered the detection effectiveness. The primary contributions of this research are fourfold. Firstly, a bilateral filter was applied to create a space-range adjustable learning architecture. Secondly, the YOLOv7x model, enhanced fromYOLOv7 was adopted to train the passive terahertz object detection model. Thirdly, NVIDIA Triton Inference Server is used to deploy the model to test the real-time detection capabilities. Lastly, a client code was developed to perform remote inference via HTTP connection. The proposed method was evaluated using passive terahertz images collected from different scenarios. The results indicate that the bilateral filter technique significantly improves the accuracy of detection of all experimented models on passive terahertz images. Among these, the model YOLOv7x_filtered_2_25 demonstrated the best performance in terms of real-time capability and detection accuracy. The single-class and two-class detection accuracies reached 93.63% and 96.44%, respectively, and the average HTTP latency of our YOLOv7x_filtered_2_25 model deployed on Triton Inference Server is around 36.7ms, and the throughput of our model is about 27.2 infer/sec.