Source-linked AI summary
TokenComSR: Task-Sensitivity-Guided Token Communication for Wireless Image Super-Resolution
Ye Wang, Li Qiao, Zhen Gao, Hua Wang
TL;DR
Wireless edge-to-server super-resolution must handle channel distortion, while existing approaches have limited support for token-level protection and pre-decoding correction. TokenComSR combines task-sensitive token power allocation with SNR-conditioned token refinement and receiver-side SR; experiments report improved reconstruction fidelity and perceptual quality across evaluated conditions.
Problem
Fading-channel distortion limits HR reconstruction from compact LR observations, and existing JSCC-SR approaches lack explicit token-level unequal protection and targeted pre-decoding refinement.
Source: §I · p. 1Method
TokenComSR combines task-sensitivity-guided token power allocation with SNR-conditioned pre-decoding token refinement and receiver-side super-resolution.
Source: §I · p. 1Results
Experiments report improved reconstruction fidelity and perceptual quality, with robust performance across evaluated SNRs, Rician factors, channel-use budgets, datasets, and SR scales.
Source: §V · p. 6Takeaways & Limitations
The evaluated TokenComSR framework supports wireless transmission of LR observations followed by receiver-side HR reconstruction across the tested conditions.
Source: §V · p. 6
Abstract
from arXiv · showhide
For resource-constrained wireless edge devices over bandwidth-limited fading channels, wireless image transmission using traditional separate coding suffers from the cliff-effect collapse. Prevailing deep joint source-channel coding (JSCC) based on convolutional neural networks can mitigate this issue but usually fail to preserve patch-level structures, thereby preventing adaptive per-token power allocation and limiting token-domain compensation for super-resolution (SR). To address these challenges, we propose a token communication framework with SR (TokenComSR). Specifically, we conceive a task-sensitive power allocation (TSPA) module and a signal-to-noise ratio (SNR)-conditioned token refinement module (TRM). TSPA distills training estimates of task sensitivity into inference token power weights, while TRM estimates an SNR-conditioned residual to correct channel-induced distortion in the token domain before decoding. Building on TSPA and TRM, the proposed TokenComSR pairs a Swin Transformer-based token transceiver with a receiver-side SR module for resource-constrained wireless image transmission. Simulation results confirm the effectiveness of the proposed TSPA and TRM, demonstrating improvements over separate coding and JSCC-SR baselines in both reconstruction fidelity and perceptual quality.
I. INTRODUCTION
The introduction motivates token-level communication for edge-to-server super-resolution, where channel distortion and CNN-based feature entanglement limit reconstruction and adaptive protection. TokenComSR combines task-sensitive power allocation, pre-decoding token refinement, and progressive training.
I. INTRODUCTION
- Resource-constrained edge devices offload compact LR observations for HR processing, but fading-channel distortion limits output quality through the reconstruction pipeline.Separate coding can suffer abrupt quality collapse below a channel-SNR threshold, while deep JSCC mitigates that cliff effect.Source: §I · p. 1
- CNN-based JSCC-SR and adaptive JSCC lack explicit token-level unequal protection and targeted pre-decoding refinement, despite regions contributing unequally to HR reconstruction.Token-patch correspondence offers a structured basis for per-token allocation and correction.Source: §I · p. 1
- TokenComSR integrates JSCC with learned SR, using TSPA at the transmitter and SNR-conditioned TRM before receiver-side image decoding.The framework exploits token-patch correspondence and includes progressive training for end-to-end optimization.Source: §I · p. 1
- TSPA distills downstream task sensitivity into unequal token power under fixed channel-use and average-power constraints, while TRM repairs noisy tokens before decoding and SR.A progressive four-phase strategy coordinates the cascaded pipeline through staged activation and dedicated loss combinations for stable end-to-end convergence.Source: §I · p. 1
II. SYSTEM MODEL
The system encodes an LR image as latent tokens, allocates token power, and transmits them within a fixed channel-use budget. At the receiver, token refinement precedes decoding and SR reconstruction over a quasi-static fading channel.
II. SYSTEM MODEL
- The transmitter maps the LR observation into N latent tokens, applies token-level power allocation and average-power normalization, then packs real-valued entries into complex channel symbols.The receiver refines the received tokens before cascaded image decoding and SR reconstruction.Source: §II · p. 2, Figure 1 · p. 2
- The channel-use ratio is defined relative to the complete HR image, and CBR denotes this HR-relative quantity unless otherwise stated.This definition measures channel usage against the RGB samples in the target image.Source: §II · p. 2
- The considered uplink uses quasi-static Rician block fading, with one channel realization shared across an image and independently resampled across images.K = 0 is the Rayleigh limit; TSPA allocates unequal protection under common channel attenuation, not token-selective fading.Source: §II · p. 2
A. Task-Sensitive Power Allocation (TSPA)
TSPA predicts token-wise scores and uses task-sensitivity estimates to form inference-time power coefficients. Its training teacher derives allocation targets from first-order sensitivity under a fixed total-power budget.
A. Task-Sensitive Power Allocation (TSPA)
- TSPA uses a lightweight token-wise MLP to produce a positive score for each encoder token.The predictor uses layer normalization, a bottleneck linear layer, GELU activation, and a scalar output layer.Source: A. Task-Sensitive Power Allocation… · p. 2
- During training, a task-sensitivity teacher uses first-order token sensitivity to the HR-domain L1 loss to supervise the allocation predictor.The teacher weights are defined under a first-order surrogate with a fixed total-power budget.Source: A. Task-Sensitive Power Allocation… · p. 3
- At inference, the teacher is removed and predicted scores become mean-one power coefficients that emphasize structurally informative regions.The nonuniform allocation also retains protection for reconstruction-relevant non-edge regions.Source: A. Task-Sensitive Power Allocation… · p. 3
B. Token Refinement Module
TRM refines received tokens before image decoding by estimating an SNR-conditioned residual. Its Transformer blocks use global self-attention to restore cross-token relationships disturbed by the channel.
B. Token Refinement Module
- TRM estimates an SNR-conditioned residual in token space before image decoding, rather than leaving channel distortion entirely to the decoder.The motivation is that decoding and subsequent SR may amplify remaining errors.Source: B. Token Refinement Module · p. 3
- TRM projects noisy received tokens into a refinement space and processes them with stacked Transformer refinement blocks conditioned on nominal SNR.The implementation uses two refinement blocks.Source: B. Token Refinement Module · p. 3
- Zero-initialized residual gates make the initial TRM mapping approximate identity, while global self-attention uses cross-token context to restore channel-disturbed relationships.TRM is optimized under the final SR task objective, not direct supervision from clean latent tokens.Source: B. Token Refinement Module · p. 3
C. SR Reconstruction Module in TokenComSR
The receiver-side SR module processes decoded LR features to synthesize missing high-frequency detail and produce the HR output. The supplied table and figure captions identify generalization and visual-comparison evaluations.
C. SR Reconstruction Module in TokenComSR
- The SR module applies shallow feature extraction, stacked residual Swin-T blocks, and sub-pixel convolution upsampling to reconstruct the HR output.The residual blocks exploit preserved structural information to synthesize missing high-frequency detail; pixel shuffle expands spatial dimensions by the upscaling factor.Source: C. SR Reconstruction Module in… · pp. 3–4
- Table I covers four-model generalization across ×2 and ×4 SR benchmarks at matched channel usage, while Fig. 4 compares visual outputs at 10 dB.The supplied captions do not state the specific benchmark outcomes or visual differences.Source: Table I · p. 4, Figure 4 · p. 4
- The residual Swin-T blocks combine local spatial features with global Transformer interactions before pixel-shuffle upsampling.The deep features are mapped to the target HR output.Source: C. SR Reconstruction Module in… · pp. 3–4
D. Progressive Training
Progressive training coordinates the cascaded pipeline through phase-specific active parameters and loss weights. It moves from encoder and predictor training to joint fine-tuning of all modules.
D. Progressive Training
- Each training phase optimizes its active parameters using weighted LR and HR pixelwise L1 losses, HR VGG-feature loss, and task-sensitivity loss.The expectation is over training images and channel realizations; the mean-one power constraint removes the need for an additional power penalty.Source: D. Progressive Training · p. 4
- Phase 1 trains the encoder, decoder, and TSPA predictor with TRM bypassed and SR frozen; Phase 2 trains the decoder, TRM, and predictor with encoder and SR frozen.Phase 3 additionally unfreezes SR.Source: D. Progressive Training · p. 4
- Phase 4 restores the best Phase-3 model and fine-tunes all modules with reduced learning rates; task-sensitivity weighting is annealed to zero during Phase 2.The task-sensitivity weight remains zero in subsequent phases.Source: D. Progressive Training · p. 4
IV. EXPERIMENTAL RESULTS
The experiments compare TokenComSR with learning-based JSCC systems under shared evaluation conditions, using DIV2K training and several channel-use and fading settings. The section also defines the baselines, metrics, and ablation evaluation.
IV. EXPERIMENTAL RESULTS
- TokenComSR is evaluated against JSCC schemes with cascaded SR and recent JSCC-SR systems in numerical simulations.For each channel-use budget, a dedicated model is trained; inference latency is measured on NVIDIA RTX 5090.Source: §IV · p. 4, A. Experimental Settings · p. 5
A. Experimental Settings
- Training uses DIV2K with 650/75/75 images, random 64 × 64 LR crops and 128 × 128 HR targets, sampling ρ from dB and K from {0, 2, 4, 8}.Evaluation uses the same four Rician factors.Source: A. Experimental Settings · pp. 4–5
- Fig. 5 compares power-allocation strategies and TRM configurations using PSNR, SSIM, and LPIPS at CBR = 1/8.Error bars show sample standard deviation over five paired channel trials.Source: Figure 5 · p. 5
- Baselines are PRAF-JSCC, SwinJSCC-LR, and SwinJSCC-HR; PSNR, SSIM, and LPIPS are evaluated under identical conditions across multiple CBRs and Rician K-factors.The study omits separate source-channel coding and focuses on learning-based JSCC baselines.Source: A. Experimental Settings · p. 5
B. Performance Evaluation
TokenComSR leads the compared methods across the evaluated SNR range, with strong results at 10 dB and across channel-use budgets. Its results also vary little under moderate nominal-SNR conditioning errors.
B. Performance Evaluation
- Across methods, performance improves smoothly with SNR; for TokenComSR, increasing K from 0 to 8 improves PSNR and SSIM and reduces LPIPS.The evaluations use CBR = 1/8; the K trend reflects reduced fading severity as the LoS component strengthens.Source: B. Performance Evaluation · p. 5
- At 10 dB, TokenComSR achieves the highest PSNR and SSIM and lowest LPIPS across the evaluated SNR range, reaching 28.2305 dB PSNR, 0.85489 SSIM, and 0.08715 LPIPS.Relative to SwinJSCC-LR at 10 dB, it improves PSNR by 3.10 dB and SSIM by 15.66%, while reducing LPIPS by 57.6%; it also exceeds SwinJSCC-HR and PRAF JSCC in PSNR by 1.68 dB and 1.56 dB, respectively.Source: B. Performance Evaluation · p. 5
- At an actual SNR of 10 dB, conditioning offsets within ±3 dB produce only 0.0270 dB total PSNR variation and at most 0.0143 dB degradation relative to matched conditioning.The passage characterizes this as low sensitivity to moderate nominal-SNR errors.Source: B. Performance Evaluation · p. 5
- At CBRs of 1/16, 1/8, and 1/4, TokenComSR achieves 26.90, 28.23, and 29.34 dB PSNR, respectively, beating the best baseline by 1.30–1.68 dB.Advantages are also observed in SSIM and LPIPS across the three budgets.Source: B. Performance Evaluation · p. 5
C. Cross-Dataset Generalization at Different SR Scales
TokenComSR generalizes across the evaluated datasets and SR scales, while ablations show benefits from combining TSPA and TRM. The paper also reports its computational trade-offs and concludes with robustness across the tested conditions.
C. Cross-Dataset Generalization at Different SR Scales
- At ×2, TokenComSR ranks first in all three metrics across five benchmarks; at ×4, it achieves best PSNR and SSIM on four datasets and best LPIPS on three.At ×4 it also outperforms SwinJSCC-HR in all three dataset-averaged metrics; the evaluation uses five datasets without dataset-specific retraining.Source: C. Cross-Dataset Generalization at… · pp. 5–6
D. Ablation Study and Complexity Analysis
- The complete TokenComSR has the best overall reconstruction fidelity; TSPA and TRM each modestly improve PSNR and SSIM over uniform allocation without TRM.Their combination reduces five-SNR average LPIPS by 0.0085, with TRM alone contributing a 0.0057 reduction; with TRM enabled, TSPA outperforms the teacher-free allocator on all three metrics.Source: D. Ablation Study and Complexity… · p. 6
- TokenComSR uses fewer parameters than both SwinJSCC baselines, has arithmetic complexity comparable to SwinJSCC-LR and lower than SwinJSCC-HR, and has approximately one-third higher end-to-end latency.Most computation remains at the receiver, consistent with the intended compute-asymmetric edge-to-server architecture.Source: D. Ablation Study and Complexity… · p. 6
V. CONCLUSION
- The four-phase progressive training strategy coordinates the transceiver, token refinement, and SR modules; experiments show robust performance across evaluated SNRs, Rician factors, channel-use budgets, datasets, and SR scales.The conclusion identifies time-selective fading, channel-estimation errors, and variable-resolution inference as future work.Source: §V · p. 6