TDJSCC: Low Complexity and Bandwidth Efficient Deep Joint Source-Channel Coding With OFDM

The practical deployment of deep joint source-channel coding (DJSCC) in edge devices faces two critical limitations. First, the prohibitive computational complexity of deep neural networks hinders efficiency. Second, existing OFDM-based systems suffer from bandwidth inefficiency due to dedicated pilot symbol allocation. To address these challenges, we propose tensorized deep joint source-channel coding (TDJSCC), a novel DJSCC framework that integrates a tensorized convolutional neural network (TCNN) with in-band pilot-augmented orthogonal frequency-division multiplexing (OFDM). The TCNN decomposes high-dimensional convolution kernels into cascaded low-rank tensor operations through singular value decomposition (SVD). Simultaneously, our in-band pilot design eliminates dedicated pilot symbols by strategically replacing data subcarriers with pilot tones. This approach achieves 100% bandwidth efficiency while maintaining channel estimation accuracy through optimized discrete Fourier transform (DFT) interpolation. The simulation results demonstrate that the proposed TDJSCC model outperforms the existing DJSCC model on low-resolution datasets and achieves comparable performance for high-resolution datasets, with 87% fewer parameters and $3.1\times $ floating-point operations (FLOPs) reduction. Furthermore, the proposed TDJSCC achieves improved performance, significantly lower computational complexity, and full bandwidth efficiency.