Deep learning based classification in through the wall microwave imaging

Abstract

This dissertation explores the usage of artificial intelligence methodologies in microwave systems for object detection and classification, with a particular emphasis on through-the-wall imaging and subsurface threat identification. Conventional microwave imaging techniques typically depend on handcrafted feature extraction, model-based inversion, or expert-driven interpretation of measurement data. While these approaches have demonstrated effectiveness in controlled scenarios, their performance often degrades in complex environments involving material heterogeneity, clutter, multipath propagation, and noise. These limitations motivate the adoption of data-driven learning approaches capable of automatically extracting robust and discriminative features from raw or minimally processed microwave measurements. The first part of this study focuses on shape and object detection in through-the-wall microwave imaging using the YOLOv8 object detection framework. YOLOv8 is a real-time detection model and is selected due to its high detection accuracy. The network employs an EfficientNet-based backbone architecture, which enables improved feature representation while maintaining resource efficiency. To enhance robustness against scale variations, the model is trained using a multi-scale training strategy in which input images are resized to different spatial resolutions. Additionally, online data augmentation techniques are applied during training to prevent overfitting and improve generalization capability. YOLOv8n-seg model is utilized to reduce training time while preserving fast inference performance. Experimental results demonstrate consistent convergence of detection, segmentation, and classification loss components in the dataset. The second part of the thesis addresses the problem of subsurface threat detection and classification using experimentally measured microwave S-parameter data. Buried object detection has a critical importance in applications such as defense, security screening, and humanitarian demining, where accurate and reliable identification of threats is essential. In this work, a deep neural network is designed to process 21 x 21 S-parameter matrices. The proposed model classifies each measurement sample into three categories: safe objects, plastic-based threats, and metal-based threats. By learning directly from real experimental data, the model effectively captures the complex relationship between microwave scattering characteristics and object material properties. The proposed approach achieves high performance in accuracy, precision, recall, and F1 score, which demonstrates consistent and reliable classification performance without relying on simulation-generated data or manually engineered features. In summary, this thesis demonstrates that deep learning-based approaches significantly improve the detection and classification capabilities of microwave systems in both through-the-wall and subsurface sensing scenarios. By leveraging real measurement data and modern neural network architectures, the proposed methods enhance robustness, scalability, and practical applicability. The findings of this study contribute toward the development of reliable, data-driven microwave sensing systems and provide a foundation for future research in security, defense, and advanced subsurface imaging applications.