On Splitting Lightweight Semantic Image Segmentation for Wireless Communications

Ebrahim Abu-Helalah; Jordi Perez-Romero; Jordi Serra

REVIEW 3 major objections 7 minor 20 references

On Splitting Lightweight Semantic Image Segmentation for Wireless Communications

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Splitting a semantic segmentation network across a wireless link cuts required bit rate by 72.6% while keeping segmentation accuracy.

desk verdict Sensible split-inference design, but the headline bandwidth saving is an artifact of comparing an uncompressed label map with a raw feature tensor. read the letter →

arxiv 2507.14199 v1 pith:WHYBGWSR submitted 2025-07-14 cs.NI cs.CVeess.IV

Ebrahim Abu-Helalah , Jordi Serra , Jordi Perez-Romero This is my paper · ORCID

classification cs.NIcs.CVeess.IV

keywords semanticcommunicationimagesegmentationmodelsplittingPIDNetbitratereductionGPUutilizationAWGNchannel6Gnetworks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes a semantic communication scheme in which a lightweight image-segmentation network (PIDNet) is split between a resource-constrained transmitter and a more powerful receiver. The transmitter executes only the first five stages and sends a 16x16 feature tensor; the receiver runs the final pyramid-pooling and convolution stages to produce the 128x128 segmentation. In simulations over an AWGN channel with QPSK and 16QAM, this lowers the required bit rate by 72.6% compared with sending the full segmentation output from the transmitter, cuts transmitter GPU processing utilization by 19.8%, and achieves the same mIoU as full transmitter-side segmentation even at 1-1.5 dB lower SNR. The aim is to make semantic image segmentation practical for bandwidth- and computation-limited devices such as mobiles and drones.

What carries the argument

The load-bearing object is PIDNet's six-stage encoder, split after stage 5. Stage 5 outputs a 16x16-pixel feature map that concentrates semantic information at coarse scale; stage 6 normally applies a Pyramid Pooling Module and convolutional layers to create the 128x128 output. By locating stage 6 at the receiver, the transmitted object becomes a tiny feature tensor rather than a full segmentation map, which is what yields the bandwidth saving while the receiver-side layers supply the contextual pooling needed for accuracy.

What would settle it

Run the proposed split with a stated bit-width quantizer on the 16x16 feature tensor and compare it against a full segmentation output compressed with run-length or entropy coding, both over the same AWGN channel; if the coded full map is no larger than the raw feature tensor or the feature tensor needs more than the assumed bits per value, the 72.6% saving will not survive.

Watch

Extended reading notes

Core claim

The central discovery is that splitting PIDNet after stage 5 preserves task-relevant semantics while drastically shrinking what must be sent: instead of transmitting the 128x128-pixel segmentation output, the transmitter sends a low-resolution 16x16 feature map that still allows the receiver's remaining layers (stage 6's pyramid pooling module plus convolutional layers) to reconstruct the segmentation at the same accuracy. The paper reports this in simulations on 500 Cityscapes images with QPSK and 16QAM over an AWGN channel, where the split scheme matched the mIoU of full transmitter-side segmentation while using 72.6% less bit rate and 19.8% less GPU processing at the transmitter.

Load-bearing premise

The reported bit-rate saving assumes both the 128x128 segmentation map and the 16x16 feature tensor are sent raw, without compression, quantization, or channel coding; compressing the segmentation map, or needing more bits per feature value, would change the 72.6 percent figure.

Editorial extensions

If this is right

A device with limited compute and battery can offload the expensive pyramid-pooling stage to an edge server while still controlling what is sent over the channel.
The same segmentation accuracy as full transmitter-side processing is reached at 1-1.5 dB lower SNR, so the split scheme tolerates worse channels for the same quality.
Required bit rate falls by 72.6% versus transmitting the segmentation output, and by 91% versus transmitting the raw image and segmenting at the receiver.
The split point is practical because the three-branch structure before stage 5 is harder to separate, while stage 5's single 16x16 output gives a clean interface for the radio link.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the comparison used properly compressed representations of both objects, the reported percentages could shift; compression would likely favor the full segmentation map more than the feature tensor, so the true bandwidth gain in a real codec is an open question the paper does not address.
The split point could be made adaptive: under good channels the transmitter could send more feature stages, and under poor channels fewer, trading bandwidth against accuracy dynamically.
The same scheme should transfer to any segmentation network with a spatial bottleneck layer, not just PIDNet, as long as the receiver-side layers can reconstruct the output from the bottleneck features.
The paper's channel model is AWGN only; in fast fading or dispersive channels, the accuracy-versus-SNR comparison may need additional pilots or error protection, which would add bits to the split scheme.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The paper is a straightforward application of split inference to semantic image segmentation: cut PIDNet after stage 5, send the 16x16 feature tensor, run the last stage at the receiver. That is a sensible engineering idea, and the paper deserves credit for actually building a simulator, using Cityscapes, and reporting mIoU versus SNR for two modulations. The GPU savings (19.8% processing, 3.4% memory) are plausible given that stage 6 and the final conv layers move to the receiver.

The soft spot is exactly where the stress-test puts it: the bit-rate comparison is apples-to-oranges. The full-segmentation baseline sends a 128x128 class-label map; the split scheme sends a 16x16 feature tensor. The paper counts both as raw bits with no quantization, source coding, or channel coding. A segmentation label map is low-entropy and compresses to a small fraction of its raw size; the feature tensor is continuous, likely floats, and may have many channels. The paper's own numbers show the saving is 3.65x, not 64x, which tells you the feature tensor costs far more per element. With a fair bit budget, the 72.6% saving could shrink dramatically, possibly disappear. That is a load-bearing claim, so the comparison needs to be redone.

Other weaknesses are secondary: the GPU utilization numbers are asserted without a measurement methodology, the mIoU curves have no error bars, and the simulator is not released. The authors do acknowledge the simplified channel model.

Still, the core idea is sound: transmitting a smaller intermediate feature reduces the required bit rate, and moving compute to the receiver reduces transmitter load. Nothing about the paper is incoherent or circular. It just needs a more careful evaluation before its quantitative claims can be trusted.

I'd send it to review, but with a request for a revised bit-rate comparison and proper GPU measurement. It's the kind of paper that could become a useful reference for 6G edge inference once the accounting is fixed.

Referee Report

3 major / 7 minor

Summary. The paper proposes a semantic communication scheme in which a lightweight PIDNet segmentation model is split after stage 5: a resource-constrained transmitter computes and sends a 16x16 feature tensor, while the receiver performs the remaining PPM and classification stages to produce the 128x128 segmentation output. The authors compare this split scheme against two baselines: transmitting the raw image and performing full segmentation at the receiver, and performing full segmentation at the transmitter and sending only the segmentation map. Using a Python simulator with 500 Cityscapes images over an AWGN channel with QPSK and 16QAM, they report a 72.6% bit-rate reduction relative to full transmitter-side segmentation, a 19.8% reduction in transmitter GPU processing utilization, a 3.4% reduction in GPU memory usage, and mIoU matching the full-segmentation baseline at 1-1.5 dB lower SNR. The paper's central claim is that splitting after stage 5 yields a favorable accuracy/bit-rate/computation trade-off for 6G-oriented semantic communication.

Significance. The idea of splitting a lightweight segmentation encoder at a communication-driven split point is concrete, falsifiable, and relevant to edge inference and semantic communication. The evaluation uses a standard external dataset (Cityscapes), a well-known lightweight model (PIDNet), and a sensible baseline (full segmentation at the transmitter). There is no circularity: no prediction is derived from fitted parameters, and the split point is a stated design choice. If the reported reductions withstand a fair and explicit accounting of transmitted bits and computational measurement, the paper would be a useful contribution. However, the current quantitative claims are not yet supported because the bit-rate metric is not formally defined and the GPU measurement methodology is absent; the value of the contribution therefore depends on completing those missing parts of the evaluation.

major comments (3)

[Section IV, Fig. 2] The headline bit-rate reduction is not supported because the bit count is never defined. The 'Required bit rate' metric is described only as depending on 'the amount of information sent' and the modulation, with no equation, no bit-per-element precision, and no statement of whether source or channel coding is included. The full-segmentation scheme transmits a 128x128 class-label map, while the split scheme transmits a 16x16 stage-5 feature tensor; these are not commensurable objects. A class-label map is low-entropy and highly compressible with standard run-length or arithmetic coding, whereas a raw feature tensor is typically continuous and requires many bits per value, possibly across multiple channels. With the reported numbers (5.66 vs 1.55 Mbps, a factor of 3.65 rather than the spatial factor of 64), the comparison is clearly based on some implicit per-element bit assumption that is never stated. A fair source-coding and quantization convention could shrink, eliminate, or reverse the claimed 72.6% saving. Please state the exact number of transmitted bits per image for each scheme, the precision/quantization of the feature tensor and the segmentation map, and whether channel coding overhead is included; if source coding is not used, justify that choice or compare over a fixed end-to-end bit budget.
[Section IV, GPU metrics] The GPU processing utilization and GPU memory utilization are reported as percentages (19.8% processing reduction, 3.4% memory reduction), but the measurement methodology is not described. No GPU model, profiling tool, measurement interval, batch size, input resolution setting, or treatment of the radio protocol stack processing is given. Without these details, the computational-resource claims cannot be reproduced or independently interpreted. Please add the profiling setup and, ideally, report raw values with confidence intervals or multiple runs.
[Section IV, Fig. 4] The mIoU comparison rests on an unspecified end-to-end transmission model. The paper states that QPSK, 16QAM, and an AWGN channel are used, but it does not specify how the transmitted tensors are mapped to symbols, whether channel coding and decoding are actually applied, or how bit errors at the receiver affect the feature tensor before the receiver-side stages. The claim that the split scheme reaches the same mIoU as full transmitter-side segmentation at 1-1.5 dB lower SNR therefore cannot be assessed from the information given. Please specify the complete transmitter/receiver chain, including any quantization, symbol mapping, and error handling, and show the bit-error behavior that produces the mIoU curves.

minor comments (7)

[Section I] The word 'Finlay' should read 'Finally'.
[Fig. 2 and Fig. 3] The legend labels 'Spitted Sem' and 'Splitted Sem' should read 'Split Sem' or 'Proposed Split'.
[Section IV, Eq. (1)] The sentence introducing the IoU formula is garbled ('calculated as and is given'); please rephrase and ensure the formula is properly defined.
[Table I] The Stage 6 row is typeset incorrectly ('P M M 128 × 1282 × Conv layers'); the operations and output size need to be readable and consistent with the text.
[Section IV, Figs. 2-4] No confidence intervals, error bars, or multiple-run statistics are reported for the bit-rate, GPU, or mIoU results, despite the use of 500 images; please indicate the variance across images or runs.
[Section V] The conclusion states 'over 90%' and '70%' reductions while the abstract and Section IV report 91% and 72.6%; make the numbers consistent after the bit-accounting issue is resolved.
[Section III, Stage 6] The text says Stage 6 creates a 128x128 feature map from a 16x16 input, but the upsampling mechanism is not explained; please clarify how the PPM and convolutional layers produce this resolution.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the main claims are design-consequence measurements and independent empirical comparisons; no fitted parameter is relabeled as a prediction.

full rationale

The paper does not fit parameters to data and then predict the same data. The split after PIDNet stage 5 is a design choice, and the reported bit-rate reduction follows structurally from transmitting a 16x16 feature tensor instead of a 128x128 segmentation output, as stated in Section III. This is a design consequence rather than a circular validation, though the exact 72.6% figure depends on unstated per-element bit counts, which is a valuation and fairness concern, not circularity. The accuracy comparison uses the external Cityscapes dataset and a pretrained PIDNet model, so the mIoU-versus-SNR results are empirical and independent of the paper's own fitted values. GPU processing and memory reductions are measured consequences of moving stage 6 to the receiver. There are no load-bearing self-citations and no imported uniqueness theorems. The only caveat is that the bit-rate metric is defined directly in terms of the amount of information sent, making the saving partly tautological, but this does not rise to a fitted-input-called-prediction or self-citation chain. The acknowledged limitation that the simulator omits dispersion and fast fading is a scope limitation, not a circular step.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central result rests on a single design choice (where to split PIDNet) and on three measurement-modeling assumptions about the channel, the GPU accounting, and the dataset. No new physical entities or fitted scalar parameters are introduced.

free parameters (1)

Split point (after stage 5) = 16x16 feature tensor
The paper chooses the cut location by hand to balance feature size against accuracy. Other split points would change bit rate, GPU load, and mIoU, and none are explored.

assumptions (4)

domain assumption AWGN channel with no dispersion or fast fading is an adequate model for the wireless link
Used in all simulations in Section IV; the conclusion explicitly acknowledges that the simulator does not fully account for these impairments.
domain assumption The simulator's GPU processing and memory utilization percentages correspond to real transmitter load
Section IV reports these as metrics but never specifies the measurement method or hardware model.
domain assumption The 500-image Cityscapes subset is sufficient to compare mIoU across schemes
Section IV states 500 images are sent per experiment; no variance or confidence intervals are reported.
domain assumption PIDNet is representative of lightweight semantic segmentation models
Section III selects PIDNet because of its balanced architecture; conclusions are drawn only from this model.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On Splitting Lightweight Semantic Image Segmentation for Wireless Communications." pith.science (2026). https://pith.science/paper/WHYBGWSR

@misc{pith2026250714199,
  author       = {Pith},
  title        = {Pith review of: On Splitting Lightweight Semantic Image Segmentation for Wireless Communications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WHYBGWSR}},
  note         = {Machine review of arXiv:2507.14199}
}

read the original abstract

Semantic communication represents a promising technique towards reducing communication costs, especially when dealing with image segmentation, but it still lacks a balance between computational efficiency and bandwidth requirements while maintaining high image segmentation accuracy, particularly in resource-limited environments and changing channel conditions. On the other hand, the more complex and larger semantic image segmentation models become, the more stressed the devices are when processing data. This paper proposes a novel approach to implementing semantic communication based on splitting the semantic image segmentation process between a resource constrained transmitter and the receiver. This allows saving bandwidth by reducing the transmitted data while maintaining the accuracy of the semantic image segmentation. Additionally, it reduces the computational requirements at the resource constrained transmitter compared to doing all the semantic image segmentation in the transmitter. The proposed approach is evaluated by means of simulation-based experiments in terms of different metrics such as computational resource usage, required bit rate and segmentation accuracy. The results when comparing the proposal with the full semantic image segmentation in the transmitter show that up to 72% of the bit rate was reduced in the transmission process. In addition, the computational load of the transmitter is reduced by more than 19%. This reflects the interest of this technique for its application in communication systems, particularly in the upcoming 6G systems.

Figures

Figures reproduced from arXiv: 2507.14199 by the authors.

**Figure 2.** Bit rate in Mbps for different experiments. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** GPU processing utilization and GPU memory utiliza [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: shows the mIoU for different executions with the two modulation schemes and for SNR values varying in the range 5-30 dB. This lets us see how the splitting process affected the accuracy of semantic segmentation. The proposed approach is able to achieve the same mIoU th…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 12 canonical work pages

[1]

Ericsson mobility report 2024,

Ericsson, “Ericsson mobility report 2024,” Ericsson AB, Tech. Rep., November 2024. [Online]. Available: https://www.ericsson.com/4adb7e/assets/local/reports-papers/mobility- report/documents/2024/ericsson-mobility-report-november-2024.pdf

work page 2024
[2]

Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,

ITU-R, “Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,” International Telecommunication Union, Tech. Rep., 2023. [Online]. Available: https://tinyurl.com/yc2bx9py

work page 2023
[3]

Goldsmith, Wireless Communications

A. Goldsmith, Wireless Communications . Cambridge, UK: Cambridge University Press, 2005. [Online]. Available: https://doi.org/10.1017/CBO9780511841224

work page doi:10.1017/cbo9780511841224 2005
[4]

A survey on edge computing systems and tools,

F. Liu, G. Tang, Y . Li, Z. Cai, X. Zhang, and T. Zhou, “A survey on edge computing systems and tools,” Proceedings of the IEEE , vol. 107, no. 8, pp. 1537–1562, 2019

work page 2019
[5]

Live uhd/4k video transmission over 5g network,

A. Lustica, M. Jurcevic, and J. Bozek, “Live uhd/4k video transmission over 5g network,” in 2021 International Symposium ELMAR , 2021, pp. 17–20

work page 2021
[6]

Vr is on the edge: How to deliver 360° videos in mobile networks

S. Mangiante, G. Klas, A. Navon, Z. GuanHua, J. Ran, and M. D. Silva, “Vr is on the edge: How to deliver 360° videos in mobile networks.” Association for Computing Machinery (ACM), 2017. [Online]. Available: https://doi.org/10.1145/3097895.3097901

work page arXiv 2017
[7]

Digital wireless semantic communication enabled by the discrete se- mantic knowledge base and updating,

Y . Shen, Z. Bao, L. Li, C. Li, H. Zhang, B. Liu, and C. Dong, “Digital wireless semantic communication enabled by the discrete se- mantic knowledge base and updating,” in 2024 IEEE/CIC International Conference on Communications in China (ICCC Workshops) , 2024, pp. 651–656

work page 2024
[8]

Semantic communications: Overview, open issues, and future research directions,

X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications: Overview, open issues, and future research directions,” IEEE Wireless Communi- cations, vol. 29, no. 1, pp. 210–219, 2022

2022

Show all 20 references

[9]

Semantic communications: Principles and challenges,

Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic communications: Principles and challenges,” 2022. [Online]. Available: https://arxiv.org/abs/2201.01389

2022 arXiv
[10]

Fully convolutional networks for semantic segmentation,

J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2015, pp. 3431–3440

2015
[11]

Pidnet: A real-time semantic segmentation network inspired by pid controllers,

J. Xu, Z. Xiong, and S. P. Bhattacharyya, “Pidnet: A real-time semantic segmentation network inspired by pid controllers,” 2023. [Online]. Available: https://arxiv.org/abs/2206.02066

2023 arXiv
[12]

Lightweight context-aware network using partial-channel transformation for real-time semantic segmentation,

M. Shi, S. Lin, Q. Yi, J. Weng, A. Luo, and Y . Zhou, “Lightweight context-aware network using partial-channel transformation for real-time semantic segmentation,” IEEE Transactions on Intelligent Transporta- tion Systems , vol. 25, no. 7, pp. 7401–7416, 2024

2024
[13]

Automatic and flexible transmission of semantic map images using polar codes for end-to-end semantic-based communication systems,

H. Rezaei, T. Sivalingam, and N. Rajatheva, “Automatic and flexible transmission of semantic map images using polar codes for end-to-end semantic-based communication systems,” in 2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Com- munication...

2023
[14]

Wireless end-to-end image transmission system using semantic communications,

M. U. Lokumarambage, V . S. S. Gowrisetty, H. Rezaei, T. Sivalingam, N. Rajatheva, and A. Fernando, “Wireless end-to-end image transmission system using semantic communications,” IEEE Access , vol. 11, pp. 37 149–37 163, 2023

2023
[15]

Pyramid scene parsing network,

H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,” 2017. [Online]. Available: https://arxiv.org/abs/1612.01105

2017 arXiv
[16]

Will 6g be semantic communications? opportunities and challenges from task oriented and secure communications to integrated sensing,

Y . E. Sagduyu, T. Erpek, A. Yener, and S. Ulukus, “Will 6g be semantic communications? opportunities and challenges from task oriented and secure communications to integrated sensing,” IEEE Network , vol. 38, no. 6, pp. 72–80, 2024

2024
[17]

Mobilenets: Efficient convolutional neural networks for mobile vision applications,

A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” 2017. [Online]. Available: https://arxiv.org/abs/1704.04861

2017 arXiv
[18]

T- yolo: Tiny vehicle detection based on yolo and multi-scale convolutional neural networks,

D. Padilla Carrasco, H. A. Rashwan, M. A. Garcisa, and D. Puig, “T- yolo: Tiny vehicle detection based on yolo and multi-scale convolutional neural networks,” IEEE Access , vol. 11, pp. 22 430–22 440, 2023

2023
[19]

The cityscapes dataset for semantic urban scene understanding,

M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” 2016. [Online]. Available: https://arxiv.org/abs/1604.01685

2016 arXiv
[20]

Generalized intersection over union: A metric and a loss for bounding box regression,

H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized intersection over union: A metric and a loss for bounding box regression,” 2019. [Online]. Available: https://arxiv.org/abs/1902.09630

2019 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Ericsson mobility report 2024,

Ericsson, “Ericsson mobility report 2024,” Ericsson AB, Tech. Rep., November 2024. [Online]. Available: https://www.ericsson.com/4adb7e/assets/local/reports-papers/mobility- report/documents/2024/ericsson-mobility-report-november-2024.pdf

work page 2024

[2] [2]

Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,

ITU-R, “Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,” International Telecommunication Union, Tech. Rep., 2023. [Online]. Available: https://tinyurl.com/yc2bx9py

work page 2023

[3] [3]

Goldsmith, Wireless Communications

A. Goldsmith, Wireless Communications . Cambridge, UK: Cambridge University Press, 2005. [Online]. Available: https://doi.org/10.1017/CBO9780511841224

work page doi:10.1017/cbo9780511841224 2005

[4] [4]

A survey on edge computing systems and tools,

F. Liu, G. Tang, Y . Li, Z. Cai, X. Zhang, and T. Zhou, “A survey on edge computing systems and tools,” Proceedings of the IEEE , vol. 107, no. 8, pp. 1537–1562, 2019

work page 2019

[5] [5]

Live uhd/4k video transmission over 5g network,

A. Lustica, M. Jurcevic, and J. Bozek, “Live uhd/4k video transmission over 5g network,” in 2021 International Symposium ELMAR , 2021, pp. 17–20

work page 2021

[6] [6]

Vr is on the edge: How to deliver 360° videos in mobile networks

S. Mangiante, G. Klas, A. Navon, Z. GuanHua, J. Ran, and M. D. Silva, “Vr is on the edge: How to deliver 360° videos in mobile networks.” Association for Computing Machinery (ACM), 2017. [Online]. Available: https://doi.org/10.1145/3097895.3097901

work page arXiv 2017

[7] [7]

Digital wireless semantic communication enabled by the discrete se- mantic knowledge base and updating,

Y . Shen, Z. Bao, L. Li, C. Li, H. Zhang, B. Liu, and C. Dong, “Digital wireless semantic communication enabled by the discrete se- mantic knowledge base and updating,” in 2024 IEEE/CIC International Conference on Communications in China (ICCC Workshops) , 2024, pp. 651–656

work page 2024

[8] [8]

Semantic communications: Overview, open issues, and future research directions,

X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications: Overview, open issues, and future research directions,” IEEE Wireless Communi- cations, vol. 29, no. 1, pp. 210–219, 2022

2022