Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Conquering High Packet-Loss Erasure: MoE Swin Transformer-Based Video Semantic Communication

T0 review · 3 major / 6 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 Video semantic codec stays usable at 90% packet loss.

desk verdict Plausible engineering system, but the headline 90%-loss numbers are computed with element-level erasure, not actual packetized transmission. read the letter →

arxiv 2508.01205 v1 pith:ML3NTLDC submitted 2025-08-02 cs.ET cs.AIcs.MM

Lei Teng , Senran Fan , Chen Dong , Haotai Liang , Zhicheng Bao , Xiaodong Xu , Rui Meng , Ping Zhang This is my paper
classification cs.ETcs.AIcs.MM
keywords semanticcommunicationpacketlossresiliencevideotransmissionMoESwinTransformerjointsource-channelcodingerasurechannelcommon-individualdecomposition
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 tries to establish that a video semantic communication system built on a MoE Swin Transformer can keep video quality usable over realistic packet-based networks, even when nine out of ten packets are dropped. It argues that semantic vectors, once interleaved, segmented, and packetized under UDP, can be reassembled with zeros in the missing places, and a receiver-side 3D CNN can reconstruct the erased semantic elements from the surviving ones using a packet-loss mask. If correct, this would let real-time video services run over UDP without retransmission, tolerate high packet loss, and coexist with existing protocol stacks. The claim is backed by simulations showing MS-SSIM above 0.6 and PSNR above 20 dB at 90% packet loss, outperforming the CNN-based MDVSC baseline.

What carries the argument

The load-bearing machinery has five parts. First, a spatial-temporal codec based on the Swin Transformer for spatial compression and the 3D Video Swin Transformer extended with a Mixture-of-Experts layer for temporal modeling, where a router picks the top-$K_r$ experts per patch. Second, a common-individual decomposition: the temporal features are projected to a common feature vector, the individual residual is downsampled further, and both are transmitted, cutting redundancy. Third, application-layer semantic interleaving with a pre-agreed random seed, which spreads correlated semantic elements across packets so a dropped packet erases scattered elements rather than a contiguous block. Fourth, the packet-loss recovery modules: 2D and 3D convolutional networks take the received incomplete semantic vectors and the packet-loss mask matrix, generate only the missing entries, and add them back. Fifth, the packetization theory, which expresses packet loss rate as a function of symbol error rate, packet length, and number of packets, and lets the system trade header overhead against robustness to meet a receiver performance threshold.

What would settle it

Run the trained decoder on UDP loss traces from a real network or a bursty loss model (for example, a Gilbert-Elliott channel) with the same average loss rate, keeping the same packetization and interleaving; if MS-SSIM and PSNR fall below the reported 0.6 and 20 dB at 90% loss, the independent element-erasure assumption that carries the training and evaluation would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that packet loss in digital semantic video transmission can be made tolerable by treating the channel as a semantic erasure channel and learning to recover erased elements. The paper's MSTVSC encoder maps a group of video frames into a semantic information matrix; after application-layer random interleaving and segmentation, each segment becomes a packet under UDP, and any packet with a symbol error is discarded. The receiver de-interleaves, replaces missing elements with zeros, and feeds the incomplete semantic matrix and a mask matrix into 3D CNN recovery modules that predict the missing entries from the un-lost ones. The paper further derives the packet loss rate as $P_L = 1 - (1 - P_s)^{8(L_{\text{head}} + L_{\text{origin}}/N_p)/M}$, showing that more, shorter packets lower the loss rate at the cost of more header overhead, and combines this with a fitted performance curve to choose packetization parameters that meet a semantic quality threshold with minimal total transmission volume. Two-stage training first learns clean reconstruction, then fixes the encoder and trains the decoder and recovery modules on randomly erased semantic elements, producing the reported robustness at 90% packet loss.

Load-bearing premise

The system's training and evaluation model packet loss as independent random erasure of individual semantic elements with a known drop probability, whereas real networks drop packets in bursts at the packet or segment level, and this approximation is not validated against real loss traces.

Editorial extensions

If this is right

  • UDP-based real-time video can remain viewable at very high packet loss without retransmission, avoiding the latency that TCP retransmission adds.
  • Packetization parameters (packet length and number of packets) can be chosen analytically to meet a target semantic quality with minimal header overhead under a given symbol error rate.
  • The common-individual decomposition and individual downsampling reduce the bandwidth ratio while preserving fast-moving detail, as shown by the ablation results.
  • Freezing the encoder in Stage II training preserves clean-channel quality while the decoder and recovery modules learn to fill erased elements, so robustness is added without sacrificing low-loss performance.
  • The approach is compatible with existing UDP/IP/Ethernet stacks, so it could be deployed without changing the network layer.

Reading between the lines

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

  • Beyond the paper: the element-wise erasure model is optimistic for real networks, since bursts of consecutive packet drops can erase correlated semantic elements even after interleaving; testing against bursty loss traces would show how much margin the recovery modules actually have.
  • Beyond the paper: the mask-matrix recovery idea is not video-specific; the same common-individual decomposition plus mask-conditioned 3D CNN could be applied to other packetized semantic modalities such as point clouds or volumetric video.
  • Beyond the paper: the theoretical packetization analysis could be inverted into a rate-robustness optimization: given a channel symbol error rate and a quality target, the system could adaptively choose packet length, number of packets, and quantization bits online.
  • Beyond the paper: the pre-agreed random seed for interleaving doubles as a cheap encryption layer, but it also creates a synchronization failure mode: if the seed or segment labels are lost, the whole GOP becomes undecodable; the paper does not analyze this failure mode.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes MSTVSC, a video semantic communication system designed to operate over packet-based UDP-style channels with packet loss. The transmitter encodes GOPs into semantic vectors, applies semantic-level interleaving and segmentation, and the receiver uses a 3D CNN recovery module with a packet-loss mask to reconstruct missing semantic elements. The paper derives a packet-loss probability expression from symbol error rate, packet length, and number of packets, and uses an empirically fitted semantic-performance curve to select communication parameters. Experiments compare MSTVSC with MDVSC and H.264/H.265-based schemes, and report MS-SSIM above 0.6 and PSNR above 20 dB at 90% packet loss.

Significance. If the reported performance holds for actual packet loss, the result would be a practically relevant contribution: video communication that remains usable at very high packet loss without retransmission, using off-the-shelf upper-layer protocols. The paper contains a useful derivation of packet-loss probability under independent symbol errors, a thoughtfully staged training procedure to preserve quality at low loss while learning recovery at high loss, and a clearly described architecture with interleaving, common/individual feature separation, MoE-based 3D Swin Transformer, and a mask-conditioned recovery module. The ablation study isolates the contributions of feature extraction, packet-loss recovery, and MoE. However, the central quantitative claim is tied to evaluation under element-level independent erasure rather than real packet-level loss, so the headline numbers are not yet supported as packet-loss results.

major comments (3)
  1. [Sec. V-A.4 and Fig. 11] The training and evaluation protocol models packet loss as independent per-element erasure: 'each element in the semantic information matrix may be dropped with a given packet loss rate PL' (Sec. V-A.4), and Fig. 11 is described as performance 'under a packet loss channel.' This is not the same as the packet loss defined in Eq. (1), where an entire packet is discarded. After the application-layer segmentation and interleaving described in Sec. II, losing a packet removes all elements in that segment; after de-interleaving, the missing set is a without-replacement subset of the transmitted elements, not an independent Bernoulli mask. The two models differ in the variance of the number of missing elements and in higher-order correlation structure, which can materially affect a correlation-based recovery module. No packet-level simulation, bursty-loss model, or UDP/IP/Ethernet test is reported. The claim of MS-SSIM > 0.6 and PSNR > 20 dB at 90% packet loss should be re-evaluated with actual packetization and segment-level loss, or the paper should clearly state that the results are for an element-level semantic erasure channel and not for packet loss.
  2. [Sec. III, Eqs. (8)-(10) and Fig. 10] The optimization result in Eq. (10) uses fp, a semantic-performance curve fitted from the authors' own model evaluated under the element-level erasure model. Because fp is measured from MSTVSC itself, Eq. (10) is a calibration identity rather than an independent predictive derivation: any monotone decreasing fp automatically yields an Np threshold once the curve is known. The text calls this a 'theoretical derivation' and uses it to draw optimal-parameter planes in Fig. 10, but no validation is given that this calibration transfers to the packet-level channel defined in Eq. (1). To make the claim load-bearing, the authors should either derive an analytical or at least a model-based relation between PL and semantic performance that is not fit to the test curve, or validate the fitted fp on a held-out packet-level simulation.
  3. [Sec. III, Fig. 9, and Sec. V-B.1] The theoretical packet-loss rate in Eq. (6) is derived under the assumption that symbol errors are independent across the symbols of every packet and that all packet loss arises from error-driven discard. This is a reasonable idealized model, but the paper does not simulate the physical-layer modulation and the complete protocol stack to verify the mapping from SER to packet loss and then to semantic performance. The surfaces in Figs. 9 and 10 combine Eq. (6) with the element-level fp curve, so the end-to-end relationship between SER and semantic performance is not empirically established. A concrete end-to-end simulation with symbol errors, packetization, discard of erroneous packets, and de-interleaving would close this gap and would also test the validity of the independent-element approximation at the heart of the training procedure.
minor comments (6)
  1. [Abstract] The phrase 'an packet-loss mask matrix' should be 'a packet-loss mask matrix'.
  2. [Sec. III, preceding Eq. (9)] The text refers to 'the derivative of Equation (23)', but no Eq. (23) exists in the paper; this should be Eq. (6).
  3. [Sec. V-B.1] The heading contains a typo: 'The V ariation' should be 'The Variation'.
  4. [Fig. 10 caption] The caption reads 'with SER, packet length and packet length'; the second occurrence should presumably be 'number of packets'.
  5. [Sec. V-A.4] The training packet-loss array [0.0, 0.3, 0.6, 0.9] is used to select drop rates per GOP, but no details are given on how the element-drop probability relates to the packet-loss probability in Eq. (1); this connection should be stated explicitly even in the revised evaluation.
  6. [Sec. V-B.4] The abbreviation 'FE' is used in the ablation discussion ('MSTVSC without the feature extraction (FE) system') but is not defined earlier; define it at first use.

Circularity Check

1 steps flagged · score 4.0 of 10

Section III's packet-length 'derivation' composes the paper's own fitted fp curve with Eq. (1); the central MSTVSC-vs-MDVSC comparison remains independent, so circularity is partial.

  1. fitted input called prediction [Section III (Theoretical Derivation), paragraph beginning 'By fitting the experimental relationship curve fp...' through Eq. (10); reflected in Fig. 10.]
    "By fitting the experimental relationship curve fp between packet loss rate and semantic performance, the connection between these parameters and semantic performance is established, guiding system optimization. Substituting PL into fp enables the relationship between semantic performance and communication parameters to be quantified... ∀ ΓP ≤ ΓMAX(Ps), ∃ Npmin s.t. fp(PL) ≥ ΓP (10)."

    The claimed 'theoretical analysis' of packetization's effect on semantic performance is built on fp, an empirical curve fitted to the authors' own MSTVSC simulation data. Equation (10) then 'derives' the existence of an optimal packet count Npmin solely from the monotonicity of this fitted curve; the semantic performance as a function of packet length is fp(PL(Np)) by construction, and Eq. (8) defines ΓMAX as fp evaluated at the minimal PL. Thus the derived 'impact of packet length on semantic performance' reduces to the fitted input composed with Eq. (1), not to an independent first-principles prediction.

full rationale

The paper's central system claim is independently supported: MSTVSC is trained and evaluated on external datasets (Vimeo-90k, HEVC, UVG) and compared against MDVSC and H.264/H.265, so the reported MS-SSIM/PSNR versus packet loss curves are empirical outputs rather than consequences of the packetization derivation. No load-bearing self-citation chain appears: self-citations such as [9], [16], [18], and [30] support baselines or framing observations, but none is invoked as a uniqueness theorem that forces the architecture. The one genuine circular element is the packetization 'derivation' in Section III, where the fitted curve fp is renamed as a theoretical relationship and then used in Eq. (10) to assert the optimal packet count; this is a fitted input called a derivation. The paper also models packet loss in training as independent per-element erasure while defining PL in Eq. (1) as whole-packet loss, but that is an external-validity concern rather than a circularity, and no explicit limitation passage in the manuscript acknowledges it. Overall, the circularity is real but localized, so a score of 4 is appropriate.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard protocol assumptions and on two ad hoc modeling choices: element-level erasure as a stand-in for packet loss, and the empirical fp curve fitted from the authors' own model. No new physical or mathematical entities are introduced.

free parameters (4)
  • fp(PL), empirical semantic-performance curve
    Fitted from MSTVSC simulation results in Fig. 11 and used in Eqs. (8) and (10) to map packet loss rate to MS-SSIM/PSNR and to select packet length.
  • Semantic performance threshold Gamma_P = 0.7 MS-SSIM example
    Chosen by hand as the receiver performance requirement to demonstrate parameter optimization; not learned from data.
  • MoE expert count and top-K
    Architecture hyperparameters that affect performance but are not reported in the paper; needed to reproduce the model.
  • Loss weights alpha, beta, gamma
    Used in Eqs. (15) and (17) for the two-stage training losses; values are not reported and would be tuned on Vimeo-90k.
assumptions (5)
  • domain assumption Symbol errors within a packet are independent and each packet is discarded if it contains at least one symbol error.
    Used to derive Eq. (1) in Section III, matching typical UDP/IP error-prone packet discard behavior.
  • domain assumption The semantic performance versus packet loss rate curve fp is monotonically decreasing.
    Stated in Section III as determined from experimental simulations; used to justify the existence of Np_min in Eq. (10).
  • ad hoc to paper Element-level random erasure at rate PL is an adequate model of packet loss after semantic interleaving and segmentation.
    Training Stage II in Section V.A.4 drops individual semantic elements with probability PL and fills zeros; real packet loss is bursty and segment-structured, and the paper does not validate the equivalence.
  • ad hoc to paper Common features dominate GOP redundancy and individual features can be heavily downsampled without losing necessary detail.
    Motivates the Common and Individual Feature Extractors in Section IV.C; supported only by the ablation in Fig. 13.
  • domain assumption Pre-agreed random seed allows transmitter and receiver to share interleaving permutations without signaling.
    Used in Section II.A; assumes no seed mismatch or synchronization failure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conquering High Packet-Loss Erasure: MoE Swin Transformer-Based Video Semantic Communication." pith.science (2026). https://pith.science/paper/ML3NTLDC

@misc{pith2026250801205,
  author       = {Pith},
  title        = {Pith review of: Conquering High Packet-Loss Erasure: MoE Swin Transformer-Based Video Semantic Communication},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ML3NTLDC}},
  note         = {Machine review of arXiv:2508.01205}
}
read the original abstract

Semantic communication with joint semantic-channel coding robustly transmits diverse data modalities but faces challenges in mitigating semantic information loss due to packet drops in packet-based systems. Under current protocols, packets with errors are discarded, preventing the receiver from utilizing erroneous semantic data for robust decoding. To address this issue, a packet-loss-resistant MoE Swin Transformer-based Video Semantic Communication (MSTVSC) system is proposed in this paper. Semantic vectors are encoded by MSTVSC and transmitted through upper-layer protocol packetization. To investigate the impact of the packetization, a theoretical analysis of the packetization strategy is provided. To mitigate the semantic loss caused by packet loss, a 3D CNN at the receiver recovers missing information using un-lost semantic data and an packet-loss mask matrix. Semantic-level interleaving is employed to reduce concentrated semantic loss from packet drops. To improve compression, a common-individual decomposition approach is adopted, with downsampling applied to individual information to minimize redundancy. The model is lightweighted for practical deployment. Extensive simulations and comparisons demonstrate strong performance, achieving an MS-SSIM greater than 0.6 and a PSNR exceeding 20 dB at a 90% packet loss rate.

Figures

Figures reproduced from arXiv: 2508.01205 by the authors.

Figure 1
Figure 1. The semantic digital communication system model proposed in this paper is depicted by the black arrow in the figure. At the transmitter, source data is encoded by the semantic encoder to generate semantic data. This data is interleaved to form a disordered semantic data vector, which is then segmented into semantic data segments (SDS). An application-layer header (AH) is appended to each segment. The data subsequent… view at source ↗
Figure 2
Figure 2. A schematic diagram of application-layer semantic information element interleaving. In the semantic data on the left, different colors represent the outputs of different channels in the semantic information matrix. Initially, the semantic data is flattened into a one-dimensional vector. Subsequently, a permutation index of the same length as the semantic data is generated based on a specified random seed. Interleavi… view at source ↗
Figure 3
Figure 3. Application layer segmentation flowchart. In UDP, data segmentation is not conducted at the transport layer, so if semantic data is directly processed by UDP, any symbol error results in the entire packet being discarded, causing complete loss of semantic data. However, the part of error-free semantic information can still be used for decoding video semantic data. To address this, semantic data is seg￾mented at the … view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: The overall architecture of our MSTVSC for wireless image transmission. where PLmin (Ps) represents the minimum PL under a given Ps. Consequently, the maximum performance ΓMAX(Ps) can be expressed as: ΓMAX(Ps) = fp  1 − (1 − Ps) 8·Lhead M  . (8) Furthermore, when Lhe…
Figure 5
Figure 5. Figure 5: The architecture of the spatial semantic code. Transformer as described in [31]. Within the spatial semantic encoder, a single frame Xgk ∈ R H×W×3 from the GOP is divided into H/2 × W/2 non-overlapping patches. These patches are handled by the Patch Embedding module to…
Figure 6
Figure 6. Figure 6: The architecture of the spatial-temporal semantic codec. ing operation, reducing the concatenated embeddings from a dimension of 4C1 to C2. The resulting tokens, with a resolution of H/4 × W/4, are subsequently processed by N2 Swin Transformer blocks. Stage 3 adopts a …
Figure 7
Figure 7. Figure 7: The architecture of the Common Feature Extractor and the Individual Feature Extractor. To further compress the redundancy of common information in the temporal dimension, the Common Feature Extractor and Individual Feature Extractor are designed as shown in [PITH_FULL…
Figure 8
Figure 8. Figure 8: The architecture of the Packet loss recovery module. To address packet loss in communication, a Common Packet Loss Recovery Module and an Individual Packet Loss Re￾covery Module are developed, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: A three-dimensional surface plot illustrating the variation of packet loss rate with SER, number of packets and packet length is presented. A plane corresponding to a packet loss rate of 15% is drawn in the figure, intersecting the surface along a black line. This blac…
Figure 10
Figure 10. Figure 10: A three-dimensional surface plot illustrating the variation of semantic performance metrics MS-SSIM with SER, packet length and packet length is presented. Planes corresponding to MS-SSIM of 0.7 are drawn in the figure, intersecting the surface along black lines. Thes…
Figure 11
Figure 11. Figure 11: The relationship between MS-SSIM and PSNR performance and the packet loss rate under a packet loss channel is illustrated. function of channel conditions and communication parameters is presented in [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 12
Figure 12. Figure 12: Visualization of Common and Individual Features. and individual feature extraction, are shown in [PITH_FULL_IMAGE:figures/full_fig_p009_12.png]
Figure 13
Figure 13. Figure 13: The ablation experiment result. compression efficiency, enabling more effective semantic in￾formation compression at the same CBR. When compared to MSTVSC without the packet loss re￾covery (PLR) system, the performance of the latter declines more rapidly as packet los…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SITP: A High-Reliability Semantic Information Transport Protocol Without Retransmission for Semantic Communication

    eess.IV 2025-12 conditional novelty 3.0 of 10

    SITP is a header-only-checksum transport that keeps corrupted payloads for a learned semantic decoder, giving UDP-like latency with better reconstruction under noise and burst losses.

Reference graph

Works this paper leans on

42 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Intellicise wireless networks from semantic commu- nications: A survey, research issues, and challenges,

    P. Zhang et al. , “Intellicise wireless networks from semantic commu- nications: A survey, research issues, and challenges,” IEEE Commun. Surveys Tuts., Early Access

    work page

  2. [2]

    Recent contributions to the mathematical theory of com- munication,

    W. Weaver, “Recent contributions to the mathematical theory of com- munication,” ETC: A Review of General Semantics , vol. 10, no. 4, pp. 261–281, 1953

    work page 1953

  3. [3]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” Bell Syst. Tech. J., vol. 27, no. 3, pp. 379–423, Jul. 1948

    1948

  4. [4]

    Toward wisdom-evolutionary and primitive-concise 6G: A new paradigm of semantic communication networks,

    P. Zhang et al., “Toward wisdom-evolutionary and primitive-concise 6G: A new paradigm of semantic communication networks,” Engineering, vol. 8, pp. 60–73, Jan. 2022

    work page 2022

  5. [5]

    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 Commun. , vol. 29, no. 1, pp. 210–219, Feb. 2022

    work page 2022

  6. [6]

    A paradigm shift toward semantic communications,

    K. Niu et al., “A paradigm shift toward semantic communications,”IEEE Commun. Mag. , vol. 60, no. 11, pp. 113–119, Nov. 2022

    work page 2022

  7. [7]

    Semantic communication system based on semantic slice models propagation,

    C. Dong et al. , “Semantic communication system based on semantic slice models propagation,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 202–213, Jan. 2023

    work page 2023

  8. [8]

    Model division multiple access for semantic communi- cations,

    P. Zhang et al., “Model division multiple access for semantic communi- cations,” Front. Inf. Technol. Electron. Eng., vol. 24, no. 6, pp. 801–812, Jun. 2023

    work page 2023

Show all 42 references
  1. [9]

    MDVSC—Efficient wireless model division video se- mantic communication,

    Z. Bao et al. , “MDVSC—Efficient wireless model division video se- mantic communication,” IEEE Internet Things J. , vol. 12, no. 2, pp. 1109–1124, Jan. 2025

    2025

  2. [10]

    Deep learning enabled semantic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Trans. Signal Process. , vol. 69, pp. 2663–2675, 2021

    2021

  3. [11]

    Deep learning for joint source- channel coding of text,

    N. Farsad, M. Rao, and A. Goldsmith, “Deep learning for joint source- channel coding of text,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Calgary, AB, Canada, Apr. 2018, pp. 2326–2330

    2018

  4. [12]

    A perceptually mo- tivated approach for low-complexity speech semantic communication,

    X. Chen, J. Wang, L. Xu, J. Huang, and Z. Fei, “A perceptually mo- tivated approach for low-complexity speech semantic communication,” IEEE Internet Things J. , vol. 11, no. 12, pp. 22054–22065, Jun. 2024

    2024

  5. [13]

    Semantic communication systems for speech transmission,

    Z. Weng and Z. Qin, “Semantic communication systems for speech transmission,” IEEE J. Sel. Areas Commun. , vol. 39, no. 8, pp. 2434– 2444, Aug. 2021

    2021

  6. [14]

    Deep joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,” IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, Sep. 2019

    2019

  7. [15]

    DeepJSCC-f: Deep joint source-channel coding of images with feedback,

    D. B. Kurka and D. G ¨und¨uz, “DeepJSCC-f: Deep joint source-channel coding of images with feedback,” IEEE J. Sel. Areas Inf. Theory , vol. 1, no. 1, pp. 178–193, May 2020

    2020

  8. [16]

    A specific task-oriented semantic image communication system for substation patrol inspection,

    S. Fan et al. , “A specific task-oriented semantic image communication system for substation patrol inspection,” IEEE Trans. Power Del. , vol. 39, no. 2, pp. 835–844, Apr. 2024

    2024

  9. [17]

    Nonlinear transform source-channel coding for semantic communications,

    J. Dai et al. , “Nonlinear transform source-channel coding for semantic communications,” IEEE J. Sel. Areas Commun. , vol. 40, no. 8, pp. 2300– 2316, Aug. 2022

    2022

  10. [18]

    W AFI-VSC: Wireless adaptive frame interpolation video semantic communication,

    X. Qi, N. Ma, Z. Bao, Y . Liu, C. Dong, and X. Xu, “W AFI-VSC: Wireless adaptive frame interpolation video semantic communication,” in Proc. 16th Int. Conf. Wireless Commun. Signal Process. (WCSP) , Hefei, China, 2024, pp. 1503–1508

    2024

  11. [19]

    Agent-driven generative semantic communication with cross-modality and prediction,

    W. Yang, Z. Xiong, Y . Yuan, W. Jiang, T. Q. S. Quek, and M. Debbah, “Agent-driven generative semantic communication with cross-modality and prediction,” IEEE Trans. Wireless Commun. , vol. 24, no. 3, pp. 2233–2248, Mar. 2025

    2025

  12. [20]

    Wireless semantic communi- cations for video conferencing,

    P. Jiang, C.-K. Wen, S. Jin, and G. Y . Li, “Wireless semantic communi- cations for video conferencing,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 230–244, Jan. 2023

    2023

  13. [21]

    A semantic communication and vvc based hybrid video coding system,

    P. Samarathunga, Y . Ganearachchi, T. Fernando, A. Jayasingam, I. Alahapperuma, and A. Fernando, “A semantic communication and vvc based hybrid video coding system,” IEEE Access , vol. 12, pp. 79202– 79224, 2024

    2024

  14. [22]

    Multimodal se- mantic communication for generative audio-driven video conferencing,

    H. Tong, H. Li, H. Du, Z. Yang, C. Yin, and D. Niyato, “Multimodal se- mantic communication for generative audio-driven video conferencing,” IEEE Wireless Commun. Lett. , vol. 14, no. 1, pp. 93–97, Jan. 2025

    2025

  15. [23]

    Toward adaptive semantic communications: Efficient data transmission via online learned nonlinear transform source-channel coding,

    J. Dai et al. , “Toward adaptive semantic communications: Efficient data transmission via online learned nonlinear transform source-channel coding,” IEEE J. Sel. Areas Commun. , vol. 41, no. 8, pp. 2609–2627, Aug. 2023

    2023

  16. [24]

    DSCS: A decoupled semantic communication system for video conferencing,

    M. Shi et al. , “DSCS: A decoupled semantic communication system for video conferencing,” in Proc. 9th Int. Conf. Comput. Commun. Syst. (ICCCS), Xi’an, China, Apr. 2024, pp. 321–326

    2024

  17. [25]

    Wireless deep video semantic transmission,

    S. Wang et al. , “Wireless deep video semantic transmission,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 214–229, Jan. 2023

    2023

  18. [26]

    DeepWiVe: Deep-learning-aided wireless video transmission,

    T.-Y . Tung and D. G ¨und¨uz, “DeepWiVe: Deep-learning-aided wireless video transmission,” IEEE J. Sel. Areas Commun. , vol. 40, no. 9, pp. 2570–2583, Sep. 2022. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    2022

  19. [27]

    Y . Tian, J. Ying, Z. Qin, Y . Jin and X. Tao, ”Synchronous Multi-Modal Semantic Communication System With Packet-Level Coding,” IEEE Trans. Wireless Commun. , vol. 24, no. 5, pp. 3684-3697, May 2025

    2025

  20. [28]

    Diffusion-based wireless semantic communication for VR image,

    H. Zhang et al. , “Diffusion-based wireless semantic communication for VR image,” in Proc. IEEE/CIC Int. Conf. Commun. China (ICCC Workshops), Hangzhou, China, Aug. 2024, pp. 639–644

    2024

  21. [29]

    A semantic communication system for point cloud,

    X. Liu et al., “A semantic communication system for point cloud,” IEEE Trans. V eh. Technol., vol. 74, no. 1, pp. 894–910, Jan. 2025

    2025

  22. [30]

    sDMCM—Semantic digital modulation constellation mapping scheme for semantic communication,

    L. Teng, W. An, C. Dong, and X. Xu, “sDMCM—Semantic digital modulation constellation mapping scheme for semantic communication,” IEEE Internet Things J. , Early Access

  23. [31]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu et al. , “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , Montreal, QC, Canada, Oct. 2021, pp. 9992–10002

    2021

  24. [32]

    SwinJSCC: Taming Swin Transformer for Deep Joint Source-Channel Coding,

    K. Yang, S. Wang, J. Dai, X. Qin, K. Niu, and P. Zhang, “SwinJSCC: Taming Swin Transformer for Deep Joint Source-Channel Coding,” IEEE Trans. Cogn. Commun. Netw. , vol. 11, no. 1, pp. 90–104, Feb. 2025

    2025

  25. [33]

    Video Swin transformer,

    Z. Liu et al. , “Video Swin transformer,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , New Orleans, LA, USA, Jun. 2022, pp. 3192–3201

    2022

  26. [34]

    GShard: Scaling giant models with conditional computation and automatic sharding,

    D. Lepikhin et al. , “GShard: Scaling giant models with conditional computation and automatic sharding,” arXiv preprint , 2020. [Online]. Available: https://arxiv.org/abs/2006.16668

    2020 arXiv

  27. [35]

    Video enhance- ment with task-oriented flow,

    T. Xue, B. Chen, J. Wu, D. Wei, and W. T. Freeman, “Video enhance- ment with task-oriented flow,” Int. J. Comput. Vis. , vol. 127, no. 8, pp. 1106–1125, Aug. 2019

    2019

  28. [36]

    Common test conditions and software reference configurations,

    F. Bossen et al. , “Common test conditions and software reference configurations,” Int. Organ. Standard., Geneva, Switzerland, document JCTVC-L1100, 2013

    2013

  29. [37]

    UVG dataset: 50/120fps 4K sequences for video codec analysis and development,

    A. Mercat, M. Viitanen, and J. Vanne, “UVG dataset: 50/120fps 4K sequences for video codec analysis and development,” in Proc. 11th ACM Multimedia Syst. Conf. , Istanbul, Turkey, Jun. 2020, pp. 297–302

    2020

  30. [38]

    Overview of the H.264/A VC video coding standard,

    T. Wiegand, G. J. Sullivan, G. Bjøntegaard, and A. Luthra, “Overview of the H.264/A VC video coding standard,” IEEE Trans. Circuits Syst. Video Technol., vol. 13, no. 7, pp. 560–576, Jul. 2003

    2003

  31. [39]

    Overview of the high efficiency video coding (HEVC) standard,

    G. J. Sullivan, J.-R. Ohm, W. Han, and T. Wiegand, “Overview of the high efficiency video coding (HEVC) standard,” IEEE Trans. Circuits Syst. Video Technol., vol. 22, no. 12, pp. 1649–1668, Dec. 2012

    2012

  32. [40]

    Design of low-density parity check codes for 5G new radio,

    T. J. Richardson and S. Kudekar, “Design of low-density parity check codes for 5G new radio,” IEEE Commun. Mag., vol. 56, no. 3, pp. 28–34, Mar. 2018

    2018

  33. [41]

    Multiscale structural similarity for image quality assessment,

    Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” in Proc. 37th Asilomar Conf. Signals, Syst., Comput. , Pacific Grove, CA, USA, Nov. 2003, pp. 1398– 1402

    2003

  34. [42]

    End-to-end opti- mized image compression,

    J. Ball ´e, V . Laparra, and E. P. Simoncelli, “End-to-end opti- mized image compression,” arXiv preprint , 2016. [Online]. Available: https://arxiv.org/abs/1611.01704

    2016 arXiv

Pith tools

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