Complex Valued Linear Discriminant Analysis on mmWave Radar Face Signatures for Task-Oriented Semantic Communication
Simy M. Baby, E. S. Gopi · IEEE Transactions on Cognitive Communications and Networking · 2025
Semantic communication for classification as the intended task at the receiver involves encoding the message signal to maximize the classifier’s performance in terms of accuracy at the receiver. Most existing techniques rely on end-to-end learning, which couples data representation with the channel model, and may result in high training overhead and reduced interpretability. While decoupled pretraining separates encoder and classifier design, it often faces limitations in adapting to varying tasks and channel conditions. We propose a complex valued Linear Discriminant Analysis (CLDA) based encoding method that enhances class discrimination and feature interpretability. This method maintains decoupling and enhances robustness to channel variations by using traditional detectors like Minimum Mean Square Error (MMSE) to recover transmitted features, enabling seamless integration into existing communication systems. It is observed that the mutual information associated with classification inference using the encoded data in the proposed technique is higher than that obtained with the original feature data before encoding. This confirms the improvement of the inference in terms of accuracy. The proposed technique is first evaluated on synthetic data and then validated on the mmWave Radar Face Signatures dataset (206 classes) under a Rayleigh channel model for real-time applicability. At the receiver, we represent CLDA coefficients as Complex valued Chromatic Images (CI) to improve the performance of the Convolutional Neural Network (CNN) based classifier. Experimental results demonstrate the effectiveness of the proposed CLDA+CI framework, achieving a classification accuracy of 96.3%. The performance of the proposed method with Kernel-LDA (KLDA) is also reported for comparison.