SNR-aware Semantic Image Transmission with Deep Learning-based Channel Estimation in Fading Channels

TL;DR

This work tackles semantic image transmission over fading wireless channels by reframing the problem as end-to-end DL-based joint source-channel coding (JSCC). It introduces SwinSIT, a Swin-transformer–based encoder/decoder pair augmented with a two-phase SNR-aware semantic enhancement and a CNN-based CEAC for precise channel estimation, alongside a joint pruning-quantization scheme to meet IoT constraints. Across CIFAR10 and DIV2K-Kodak, SwinSIT demonstrates superior PSNR and MS-SSIM compared with BPG+LDPC, CNN-based Deep JSCC, and WITT, with the SNR-aware and CEAC components contributing notably to robustness in varying SNRs. The compressed variant achieves substantial size reductions (e.g., from 53.4 MB to 5.3 MB) while maintaining competitive reconstruction quality, highlighting practical viability for 6G/IoT deployments where resources are limited.

Abstract

Semantic communications (SCs) play a central role in shaping the future of the sixth generation (6G) wireless systems, which leverage rapid advances in deep learning (DL). In this regard, end-to-end optimized DL-based joint source-channel coding (JSCC) has been adopted to achieve SCs, particularly in image transmission. Utilizing vision transformers in the encoder/decoder design has enabled significant advancements in image semantic extraction, surpassing traditional convolutional neural networks (CNNs). In this paper, we propose a new JSCC paradigm for image transmission, namely Swin semantic image transmission (SwinSIT), based on the Swin transformer. The Swin transformer is employed to construct both the semantic encoder and decoder for efficient image semantic extraction and reconstruction. Inspired by the squeezing-and-excitation (SE) network, we introduce a signal-to-noise-ratio (SNR)-aware module that utilizes SNR feedback to adaptively perform a double-phase enhancement for the encoder-extracted semantic map and its noisy version at the decoder. Additionally, a CNN-based channel estimator and compensator (CEAC) module repurposes an image-denoising CNN to mitigate fading channel effects. To optimize deployment in resource-constrained IoT devices, a joint pruning and quantization scheme compresses the SwinSIT model. Simulations evaluate the SwinSIT performance against conventional benchmarks demonstrating its effectiveness. Moreover, the model's compressed version substantially reduces its size while maintaining favorable PSNR performance.

Figures (11)

• Figure 1: The model of the proposed wireless image transmission system.
• Figure 2: The proposed SwinSIT architecture.
• Figure 3: (a) The proposed SNR-aware module (b) SNR mapper block (c) excitation block.
• Figure 4: The proposed DnCNN-based channel estimator and compensator.
• Figure 5: The proposed joint pruning-quantization scheme (a) original weights (b) weights after pruning (c) quantized weights.