Dissertation Defense Announcement for Sabyasachi Biswas – 07/22/2025

Dissertation Title: “Complex-Valued Structured Parameterized Learnable Filter Banks for Time-Frequency Domain Based Classification”

Candidate: Sabyasachi Biswas

Degree: Doctor of Philosophy in Electrical & Computer Engineering

Date & Time: Tuesday, July 22, 2025, at 10:00 AM (CDT)

Location: Simrall Building, 2nd-floor Conference Room (Room 228) and via Teams [Link]

Committee: Dr. John Ball (Major Professor), Dr. Ali C. Gurbuz (Dissertation Director), Dr. Junming Diao, Dr. Ryan Green

Abstract:

Most conventional radar-based classification methods, including human activity recognition (HAR) and modulation recognition, rely on computationally intensive two-stage processes involving time-frequency (TF) transformations, such as the short-time Fourier transform (STFT), to generate micro-Doppler signature (-Ds). These -Ds are then classified using deep neural net- work (DNN), a method introducing significant temporal latency and limiting real-time applicability. To overcome these limitations, this dissertation proposes a novel complex-valued Deep learning (DL) framework employing structured parameterized learnable filter (PLF) banks, enabling direct classification from raw radar data. The research first introduces high-resolution spectrogram network (HRSpecNet), a deep learn- ing model designed to reconstruct high-resolution -Ds directly from complex-valued 1D radar data. HRSpecNet uses an autoencoder for noise suppression, a learnable STFT block for adaptive frequency transformations, and a U-Net block for High-Resolution (HR) image reconstruction. Evaluations using synthetic signals and a challenging real-world American Sign Language (ASL) dataset demonstrate improved classification accuracy by 3.48% compared to traditional STFT-based methods, highlighting HRSpecNet’s superior resolution, robustness to noise, and computational efficiency. Building on HRSpecNet, this research further introduces parameterized learnable filter net- work (PLFNet), which integrate complex-valued PLF, including Sinc, Gaussian, Gammatone, and Ricker, directly into convolutional neural network (CNN) architectures. Unlike conventional methods, PLFNet classify raw 1D radar data without explicit CNN generation, providing enhanced interpretability and computational efficiency. PLFNet achieve approximately 47% higher accuracy than standard 1D CNNs and about 7% higher accuracy than CNNs employing real-valued learnable filters. Furthermore, PLFNet match the accuracy of standard CNNs applied to μ-DS images while reducing computational latency by approximately 75%, making them particularly suitable for real-time applications. Finally, the dissertation introduces time-gated parameterized learnable filter network (TG- PLFNet), featuring time-gated parameterized learnable filters capable of adaptively focusing on critical temporal and spectral signal features, crucial for non-stationary signals. TG-PLFNet demonstrates superior performance on a newly synthesized dataset containing 51 radar and communication waveform modulations under varied conditions, surpassing existing automatic modulation recognition (AMR) models in accuracy, inference time, and interpretability. Collectively, the developed methods offer computationally efficient, interpretable, and high- performing solutions for TF-domain-based radar and radio frequency (RF) signal classification, advancing real-time capabilities and practical applicability in RF sensing applications.