Storage-Scalable Progressive Semantic Communication via Knowledge-Base Reuse

Heng Zhu; Ye Liu; Kun Zhu; Feifei Song

Source-linked AI summary

Storage-Scalable Progressive Semantic Communication via Knowledge-Base Reuse

Heng Zhu, Ye Liu, Kun Zhu, Feifei Song

arXiv:2609.10112v1Sep 9, 2026cs.LGcs.NIeess.IV

TL;DR

Existing SKBQ and MKBQ schemes trade storage against progressive refinement: SKBQ uses limited storage but has restricted capacity, while MKBQ storage grows with transmission depth. SSKBQ reuses compact KBs across residual stages and adds stage-aware supervision, achieving competitive progressive reconstruction with substantially reduced storage overhead.

Problem
Existing SKBQ has restricted quantization capacity, whereas MKBQ requires KB storage that grows linearly with transmission depth.
Method
SSKBQ reuses a compact set of KBs across multiple residual refinement stages and applies stage-aware residual supervision to regularize intermediate representations.
Results
SSKBQ achieves competitive reconstruction performance with substantially reduced storage overhead compared with existing knowledge-base quantization schemes.
Takeaways & Limitations
KB reuse provides a storage-efficient progressive semantic transmission framework while preserving progressive reconstruction capability.

Abstract

from arXiv · show

Existing knowledge-base-assisted semantic communication schemes commonly adopt either single knowledge-base quantization (SKBQ) or multi-knowledge-base residual quantization (MKBQ). SKBQ incurs limited storage overhead but has restricted quantization capacity, whereas MKBQ supports progressive refinement by assigning an independent knowledge base (KB) to each stage, causing the KB storage to grow linearly with the transmission depth. To address this problem, we propose storage-scalable knowledge-base reuse quantization (SSKBQ), which reuses a compact set of KBs across multiple residual refinement stages and thereby decouples the number of transmission stages from the number of maintained KBs. A stage-aware residual supervision mechanism is further introduced to regularize intermediate quantized representations and encourage progressive refinement. Experimental results demonstrate that KB reuse provides an effective solution to the storage scalability problem while maintaining competitive progressive reconstruction performance.

I. INTRODUCTION

Semantic communication can reduce transmission rates by sending task-relevant information, but floating-point semantic features create substantial overhead. SKBQ limits storage but refinement capacity, whereas MKBQ enables progressive reconstruction with storage that grows linearly across stages; SSKBQ reuses compact KBs to address this trade-off.

Direct transmission of extracted floating-point semantic features is costly because they are high-dimensional representations compared with compact integer inputs.The paper motivates quantization as a way to maintain transmission efficiency while preserving semantic information.
SKBQ reduces transmission rates by sending codeword indices, but its single KB restricts representation capacity and does not naturally support progressive refinement.The resulting quantization distortion can be considerable.
MKBQ supports progressive reconstruction through multi-stage residual quantization but requires an independent KB for each stage, making storage grow linearly.When each KB is large, the cumulative storage cost becomes prohibitive.
SSKBQ reuses a compact set of knowledge bases across residual refinement stages, decoupling transmission depth from KB storage growth.The scheme is designed to provide progressive reconstruction without the linear storage increase associated with assigning one KB to each stage.

II. SEMANTIC COMMUNICATION SYSTEM

The system encodes an input image into task-relevant semantic features, quantizes those features to reduce transmission overhead, and decodes the quantized representation into a reconstructed image. The design focuses on semantic-layer storage-scalable transmission under reliable delivery.

The system processes an input image with a semantic encoder to extract task-relevant features, then quantizes them before transmission and decodes them for reconstruction.The encoder and decoder form an end-to-end image reconstruction system.
The considered system abstracts physical-layer processing as a reliable delivery interface and concentrates on semantic-layer design.Channel coding, modulation, equalization, and error control are not modeled explicitly.
Quantization is introduced because directly transmitting semantic features in floating-point format incurs substantial transmission overhead.The feature map has c channels and spatial dimensions h×w.

A. Knowledge-Base Reuse Scheme

SSKBQ reuses a compact set of knowledge bases across residual refinement stages, decoupling maintained KB count from transmission depth while supporting progressive reconstruction.

SKBQ: SKBQ requires only one KB and therefore has limited storage overhead, but its restricted capacity limits reconstruction quality and prevents progressive transmission.Its reconstruction quality is often insufficient for high-resolution images.
MKBQ: MKBQ uses one independent KB per transmission stage, so storage grows linearly with transmission depth and can become prohibitively high.High-resolution reconstruction generally requires more stages, increasing cumulative storage overhead.
MKBQ: MKBQ supports progressive transmission through residual quantization, but later-stage reconstructions are not guaranteed to outperform earlier stages.This can produce unstable progressive refinement.
SSKBQ: SSKBQ addresses the storage limitation by decoupling the number of maintained KBs from transmission depth through reuse across multiple residual refinement steps.This scheme is illustrated as the proposed alternative to independent per-stage KBs.
SSKBQ: SSKBQ reuses each of K reusable KBs across assigned residual steps, supporting T progressive stages with K ≤ T maintained KBs.The first KB produces the initial approximation and subsequent residuals; later KBs are reused over their assigned refinement intervals.

B. Training Storage-Scalable Knowledge Base

The training objective combines reconstruction, quantization, and residual supervision losses to regularize intermediate representations and encourage progressively smaller residual errors.

Training objective: The overall objective combines reconstruction loss, quantization loss, and residual supervision loss in end-to-end optimization.Reconstruction loss measures distortion between the reconstructed and original images; quantization loss aligns semantic and quantized features.
Quantization loss: The quantization loss uses stop-gradient terms to update the semantic encoder toward the quantized feature and the KBs toward the semantic feature.The hyperparameter α balances these two effects.
Residual supervision: Residual supervision regularizes intermediate quantized representations because the final quantization loss alone does not constrain stage-wise refinement.The supervision explicitly targets the progressive refinement process under KB reuse.
Residual supervision: Residual supervision progressively minimizes stage-wise residual error, with 1/i assigning larger penalties to earlier refinement stages.Λ controls the overall strength of progressive supervision.

1) Datasets:

SSKBQ is evaluated on Cityscapes and COCO using standard image-reconstruction metrics under a common resized-image and progressive-transmission setup.

Datasets: Cityscapes provides urban street scenes, while COCO contains more diverse objects and visual scenes for evaluation.Cityscapes has 2,975 training images and 500 validation images; both datasets are resized to 128 × 64.
Evaluation metrics: Reconstruction quality is evaluated with PSNR, SSIM, FID, and KID.These metrics are used consistently across the evaluated datasets.
Baselines: The comparison includes JSCC as an unquantized reference, SKBQ through VQVAE variants, and MKBQ with an independent KB at each stage.The setup compares representative non-progressive and progressive schemes.
Experimental setup: The experiments use 16-stage progressive transmission with 512 codewords per KB and transmit only codeword indices during testing.Each codeword has dimension 256, and 512 spatial feature vectors are quantized at each stage.

B. Complexity and Scalability

SSKBQ reduces knowledge-base storage through reuse while retaining stage-wise transmission complexity comparable to MKBQ. Its receiver-side processing scales linearly with the number of received stages, while progressive inference latency is mainly driven by repeated decoder executions.

Storage scalability: SSKBQ requires O(KNc) storage versus O(TNc) for MKBQ, achieving only K/T of MKBQ’s KB storage when K reusable KBs serve T stages.SKBQ requires O(Nc) storage, while SSKBQ decouples maintained KB count from transmission depth.
Transmitter complexity: Nearest-neighbor assignment costs O(chwN) per stage and O(tchwN) cumulatively after t stages for both MKBQ and SSKBQ.The per-stage comparison covers hw feature vectors of dimension c against N codewords.
Receiver complexity: Receiver-side index lookup, feature assembly, and residual accumulation require O(chw) per stage and O(tchw) after t stages, independent of N.The receiver directly identifies codewords from transmitted indices.
Inference latency: Progressive inference requires O(tchw + tCdec) when decoding after every stage, so repeated decoder executions mainly determine latency despite linear KB-processing growth.Reconstruction only after the final stage instead requires O(tchw + Cdec).

C. Comparison with Existing Schemes

Experiments on Cityscapes and COCO compare SSKBQ with progressive, single-stage, and continuous-feature baselines across 16 transmission stages. SSKBQ improves over quantization baselines and reduces communication overhead, but its comparison with JSCC depends on dataset and metric.

Progressive reconstruction: Across both datasets, SSKBQ progressively improves PSNR, SSIM, FID, and KID as more transmission stages are received, and it outperforms comparable MKBQ settings.The paper attributes the MKBQ comparison to stage-aware residual supervision facilitating reused-KB optimization.
Baseline comparison: SSKBQ consistently improves over single-stage VQVAE(OH) and VQVAE(GS), while JSCC performance varies with dataset complexity and representation demands.SSKBQ achieves higher PSNR than JSCC on Cityscapes after sufficient stages, whereas JSCC performs better on COCO.
Cityscapes: 27.36 dB PSNR on Cityscapes: SSKBQ(16) improves 11.31%, 28.93%, and 18.13% over MKBQ(16), VQVAE(OH), and VQVAE(GS), respectively.The result is reported in Table I for the final reconstruction comparison.
Communication efficiency: 73,728 transmitted bits on Cityscapes are less than 1% of JSCC’s transmission overhead, while SSKBQ still improves PSNR over JSCC by 12.87%.JSCC retains better SSIM, FID, and KID on Cityscapes because it transmits continuous semantic features without quantization distortion.
COCO: 23.53 dB PSNR on COCO: SSKBQ(16) outperforms MKBQ(16), VQVAE(OH), and VQVAE(GS) by 10.16%, 24.17%, and 27.60%, respectively.JSCC nevertheless achieves higher reconstruction quality on COCO.

V. CONCLUSION

The paper addresses the storage-performance trade-off in progressive semantic communication by reusing a limited number of knowledge bases across refinement stages. Experiments show competitive reconstruction with substantially reduced storage overhead, while practical channel conditions remain future work.

SSKBQ decouples the number of maintained knowledge bases from the number of progressive transmission stages through knowledge-base reuse.
SSKBQ achieves competitive reconstruction performance with substantially reduced storage overhead compared with existing knowledge-base quantization schemes.
Increasing the number of knowledge bases does not always improve performance because semantic fragmentation and optimization difficulty may limit their effective utilization.
The study focuses on semantic-layer knowledge-base reuse under reliable index delivery, leaving noisy transmission, fading, packet loss, and channel-aware adaptation for future work.