Pith. sign in

REVIEW 4 major objections 5 minor 39 references

UAV Cognitive Semantic Communications Enabled by Knowledge Graph for Robust Object Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Knowledge-graph semantics beat image transmission for UAV detection

desk verdict A competent, incremental extension of the authors' own KG-based semantic communication work, with plausible results and honest ablations, but the SNR-robustness headline rests on an untested perfect-SNR assumption. read the letter →

arxiv 2502.03761 v1 pith:VLHCJNUF submitted 2025-02-06 eess.SP

Xi Song , Fuhui Zhou , Rui Ding , Zhibo Qu , Yihao Li , Qihui Wu , Naofal Al-Dhahir This is my paper · ORCID
classification eess.SP
keywords semanticcommunicationknowledgegraphobjectdetectionUAVSNRadaptationmulti-scalecodecjointsource-channelcodingaerialimagery
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 UAV (unmanned aerial vehicle) can send compressed semantic features instead of images and still detect objects more accurately and efficiently than conventional transmission, provided the receiver uses external knowledge to repair what noise and compression corrupt. The proposed system pairs a lightweight multi-scale feature extractor on the drone with an SNR-adaptive attention module, a multi-scale codec, and a knowledge-graph-enhanced detector on the server. Its central evidence is that on a public aerial-image benchmark this design beats both a conventional BPG+LDPC link and an attention-based deep joint source-channel coding baseline at the same channel conditions, especially at low SNR and low bandwidth ratios. If correct, this would let resource-limited drones offload detection at lower bandwidth and energy cost without sacrificing accuracy.

What carries the argument

The receiving detector's weighted graph is the load-bearing mechanism: the region-proposal network's boxes are linked to each other and to node embeddings from a large commonsense knowledge graph, with edge weights encoding classification confidence, visual similarity, and embedding distance; a relational graph attention network then aggregates the three edge types so the detector can reason about the whole scene. The SNR-adaptive module is the supporting mechanism: after each backbone block it pools the feature map, concatenates the current SNR value, and multiplies channels by an attention weight, letting one model cover many SNR regimes instead of requiring per-SNR models.

What would settle it

Run the same system on the same aerial-image benchmark with the SNR values used to compute attention weights corrupted by, say, a ±3 dB offset or zero-mean Gaussian error at test time; if mean average precision at low SNR collapses toward the no-SA baseline, then the claim that the SNR-adaptive module adapts robustly to channel conditions fails under realistic SNR estimation.

Watch

Extended reading notes

Core claim

The paper's central claim is that semantic transmission with a knowledge graph, rather than image reconstruction, is the right operating point for UAV-to-server object detection. The system extracts multi-scale semantic features on the UAV, compresses them with a parallel multi-scale codec, and transmits them over AWGN or Rayleigh channels; the server reconstructs features, not pixels. A knowledge-graph-enhanced detector then builds a weighted graph whose nodes are proposal boxes and knowledge-graph entities, aggregates it with a relational graph attention network, and uses the resulting global understanding to classify. Reported results show the system outperforms the two benchmarks under identical channel conditions, with average mAP (mean average precision) gains of 4.48% and 7.48% over the deep joint source-channel baseline at SNR = -6 dB and 5.08% and 5.63% at SNR = 9 dB in AWGN and Rayleigh channels, while using 19.3% fewer parameters and about 69% lower GFLOPs.

Load-bearing premise

The whole SNR-adaptation story depends on the UAV and the server knowing the instantaneous SNR precisely and feeding the same value into the attention computation; if real SNR estimates are noisy or delayed, the promised robustness across channel conditions is not guaranteed.

Editorial extensions

If this is right

  • The same knowledge-graph repair step should transfer to other UAV vision tasks, such as classification, tracking, and visual question answering, by building a knowledge graph for the target categories.
  • Low-SNR links can carry fewer symbols: at the lower compression ratio R = 1/9 the system still beats the baseline at R = 1/6 below 0 dB, so bandwidth can be traded for robustness.
  • A single SNR-adaptive model replaces the practice of training and switching several SNR-specific models, which cuts latency and storage on the drone.
  • Compression distortion, not channel noise, dominates at high compression ratios, so the multi-scale codec is what preserves accuracy when bandwidth is tight.

Reading between the lines

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

  • An implication the paper leaves untested is that the SNR-adaptive module assumes the SNR fed into the attention weights is accurate; adding noise or delay to that input could erase the low-SNR gains, and the degradation pattern is a direct way to test the module's sensitivity.
  • The confusion-matrix analysis suggests most residual errors are localization misses rather than misclassification, so part of the knowledge graph's gain may come from re-scoring proposals globally rather than from correcting category labels.
  • Because the knowledge graph's benefit shrinks as the bandwidth compression ratio increases, one would expect the graph to matter most in spectrum-starved deployments and to matter little once enough semantic detail survives.
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

4 major / 5 minor

Summary. This paper proposes a UAV cognitive semantic communication system for object detection. The system consists of a lightweight semantic extractor (ResNet-18 with FPN) and a multi-scale semantic encoder on the UAV, a wireless channel (AWGN or Rayleigh), and a multi-scale semantic decoder plus a knowledge-graph-enhanced object detector on a ground server. An SNR-adaptive (SA) module is introduced that feeds the instantaneous SNR into channel-attention computations at the transmitter to enable operation across SNR regimes. The detector fuses visual features with ConceptNet-derived knowledge-graph embeddings via a weighted graph and a relational graph attention network. The paper evaluates the system on the DOTA aerial dataset against BPG+LDPC and ADJSCC+Faster R-CNN benchmarks, reporting mAP improvements at low SNR (e.g., 4.48% and 7.48% over ADJSCC at -6 dB), smaller gains at high SNR, and reduced parameter count and GFLOPs. Ablation studies are provided for the SA module, knowledge graph, and multi-scale codec.

Significance. If the reported results are reliable, the paper demonstrates a promising direction: transmitting task-oriented semantic features with knowledge-graph side information can provide better detection accuracy and lower complexity than image-reconstruction-based communication for UAV-to-server object detection, particularly at low SNR and high compression. The paper is generally well structured, with explicit algorithms and multiple ablations. Particular strengths are the detailed description of the architecture, the inclusion of efficiency metrics (Table IV), and the confusion-matrix analysis of failure modes. However, the central quantitative claims are weakened by the absence of statistical error bars and by the unexamined dependence of the SA module on perfect instantaneous SNR knowledge.

major comments (4)
  1. [Section III-B / Algorithm 1 and Section IV-D] The SA module concatenates the instantaneous SNR µ with global pooled features (Algorithm 1, lines 2–4) at the UAV and, per the architecture, also at the server side. The central robustness claim in Section IV-D—that the proposed system outperforms the benchmarks especially at low SNRs—is therefore conditioned on µ being known exactly. The paper does not model SNR estimation error, feedback delay, or mismatch; the ablation in Fig. 8 only retrains the no-SA baseline at fixed SNRs and never feeds a wrong SNR to the SA module at test time. Please add experiments where the SNR input is corrupted (e.g., Gaussian error in dB or a fixed offset) and report the resulting mAP. In addition, the training SNR schedule for the full system with SA is not stated in Section IV-B, which is necessary to reproduce the reported all-in-one robustness.
  2. [Section IV, Figs. 7–10 and Table III] All experiments appear to be single training runs, with no error bars, standard deviations, or number of seeds reported. Some of the headline improvements are modest at high SNR—e.g., Table III (AWGN, R=1/6, SNR=9dB) shows AP50 +5.2% but APM −4.8%, and Fig. 7 reports 5.08% and 5.63% average improvements at SNR=9dB. Without variance estimates, it is not possible to assess whether these differences are statistically significant or robust to training randomness. Please report mean±std over at least three seeds (or equivalent confidence intervals) for the main comparisons.
  3. [Algorithm 2, lines 3–9] The loop at line 4 runs j from 0 to K−1, but line 5 indexes MCi,j where MC ∈ R^{M×C} and C is the number of detection classes (Section III-C). Since K (the number of knowledge-graph nodes, 3,579) is much larger than C (15), the indexing is out of bounds for j ≥ C. The pseudo-code appears to intend a loop over class indices (j = 0..C−1) and a mapping from class labels to knowledge-graph node embeddings. As written, Algorithm 2 is not implementable. Please correct the loops and specify the class-to-node mapping.
  4. [Section IV-B and Section IV-D] The benchmark comparison does not specify how the bandwidth compression ratio R of the semantic system maps to the BPG+LDPC system's rate (LDPC rate 1/3, 16-QAM). The text in Section IV-D claims 'transmit less data' and compares R=1/9 vs R=1/6, but it is unclear whether the transmitted data volumes per image are actually matched across systems. Please state the effective overall source-channel rate (e.g., bits per pixel or channel symbols per pixel) for each benchmark and confirm that the same or smaller rate is used for the proposed system in each comparison.
minor comments (5)
  1. [Section II, page 4] 'indpendent' should be 'independent'; similar typos occur elsewhere ('semanic', 'applid', 'multi-scle').
  2. [Introduction, reference [9]] The statement 'It was defined by Weaver and Shannon [9]' cites an opportunistic spectrum access paper, not the expected Shannon–Weaver reference; the citation should be corrected.
  3. [Section IV-D / Fig. 11] The explanation of the confusion matrix values (e.g., '0.40 means that 25% of the predictions are correct') is confusing and arithmetically unclear; please define the displayed percentages precisely.
  4. [Section IV-B] The sentence on selecting LDPC rate and modulation ('After multiple rounds of evaluation...') would benefit from reporting the parameter grid that was searched and the selection criterion.
  5. [Throughout] The spacing in 'UA V' is inconsistent (e.g., 'UA Vs' vs 'UA V'); please unify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is an empirical comparison against external benchmarks and is not derived from the paper's own assumptions or self-citations.

full rationale

The paper's central claim is an empirical performance comparison, not a mathematical derivation. The proposed system is trained on the DOTA dataset and evaluated against two external benchmarks, BPG+LDPC and ADJSCC, with matched Faster R-CNN detectors, and the reported gains are supported by ablations of the SA module, knowledge graph, and multi-scale codec (Figs. 7-10, Tables III-IV). No load-bearing step reduces to the paper's own definitions or to its self-citations: the knowledge graph is drawn from external ConceptNet, the benchmarks are external, and no uniqueness theorem or fitted parameter is imported from the authors' prior work. The citations to [1] and [16] credit the cognitive semantic communication concept and frame the paper as an extension, but the current experimental results do not depend on those citations being accepted. The SNR-adaptive module (Algorithm 1) conditions on instantaneous SNR as side information; this is an unmodeled practical assumption that limits the robustness claim, but it is not a circular reduction because the reported performance is measured rather than defined by the SNR input. The acknowledged limitation about increased misclassification for some categories is a performance caveat, not a circularity. Overall, no equation-level equivalence between inputs and claimed outputs exists, so the paper is self-contained for its empirical claims.

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

The central performance claims rest on standard deep-learning training plus several domain assumptions: perfect SNR knowledge, ConceptNet validity for DOTA categories, and AWGN/Rayleigh channel fidelity. The model weights and hyperparameters are fitted or tuned on DOTA, and no independent verification artifacts are provided.

free parameters (4)
  • Knowledge graph subgraph size = 3579 entities, 31 relation types
    Result of manual data cleaning stages (removal of isolated, low-degree, and duplicate nodes) that are not fully specified; the size and composition of the knowledge graph affect the quality of node embeddings and downstream fusion.
  • R-GAT architecture dimensions = 3 layers, hidden 512, 2 attention heads
    Hand-chosen model size with no ablation study; the information aggregation capacity of the detector depends on these values.
  • BPG+LDPC benchmark coding parameters = coding rate 1/3, 16-QAM
    Selected after multiple evaluation rounds 'in terms of the best performance'; these choices affect the fairness of the comparison and the size of the reported gains.
  • Training SNR schedule for the full system = not specified
    The SA module is trained to be SNR-adaptive, but the distribution and range of SNR values used during training are not stated; the all-SNR robustness claim depends on this schedule.
assumptions (4)
  • domain assumption ConceptNet contains relevant and correct relationships among the 15 DOTA object categories.
    Invoked in Section III.C during subgraph extraction; if ConceptNet lacks or mislabels relations for aerial objects such as harbor or roundabout, the knowledge graph module could inject noise instead of improving detection.
  • domain assumption Instantaneous SNR is known exactly at both the UAV and the server.
    The SA module concatenates SNR into the attention computation (Algorithm 1; Section III-B); no SNR estimation error is modeled, so claimed robustness under dynamic channels assumes perfect SNR knowledge.
  • domain assumption AWGN and Rayleigh fading adequately represent UAV communication channels.
    Channel model in Section II (Eqs. 1-2); real UAV links also suffer Doppler, shadowing, and interference, which are not simulated.
  • domain assumption Pretrained COCO features transfer to the DOTA aerial image domain.
    ResNet-18 with FPN pretrained on MS COCO is used as the base model (Section IV.B); if transfer is poor, the reported absolute mAP values would change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UAV Cognitive Semantic Communications Enabled by Knowledge Graph for Robust Object Detection." pith.science (2026). https://pith.science/paper/VLHCJNUF

@misc{pith2026250203761,
  author       = {Pith},
  title        = {Pith review of: UAV Cognitive Semantic Communications Enabled by Knowledge Graph for Robust Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLHCJNUF}},
  note         = {Machine review of arXiv:2502.03761}
}
read the original abstract

Unmanned aerial vehicles (UAVs) are widely used for object detection. However, the existing UAV-based object detection systems are subject to severe challenges, namely, their limited computation, energy and communication resources, which limits the achievable detection performance. To overcome these challenges, a UAV cognitive semantic communication system is proposed by exploiting a knowledge graph. Moreover, we design a multi-scale codec for semantic compression to reduce data transmission volume while guaranteeing detection performance. Considering the complexity and dynamicity of UAV communication scenarios, a signal-to-noise ratio (SNR) adaptive module with robust channel adaptation capability is introduced. Furthermore, an object detection scheme is proposed by exploiting the knowledge graph to overcome channel noise interference and compression distortion. Simulation results conducted on the practical aerial image dataset demonstrate that our proposed semantic communication system outperforms benchmark systems in terms of detection accuracy, communication robustness, and computation efficiency, especially in dealing with low bandwidth compression ratios and low SNR regimes.

Figures

Figures reproduced from arXiv: 2502.03761 by the authors.

Figure 1
Figure 1. The proposed UAV cognitive semantic communication system. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The network structure of the proposed UAV cognitive semantic communication system. In the backbone, Each “Resblock” represents a residual [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The details of the single-scale encoder, decoder, and residual [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The details of the SA module. C, H, W represent the number of channels, height, and width of the semantic features, respectively. transmitted over a noisy channel, which can be represented as Eq. (1) and Eq. (2). B. SNR Adaptation Module Existing prevalent semantic com…
Figure 6
Figure 6. Figure 6: Example of the detection scheme applied to an image. Note that in the weighted graph, pink nodes and green nodes represent visual semantic features. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: mAP versus SNR for the proposed system and two benchmarks under AWGN channel and Rayleigh channel. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: mAP versus SNR for the proposed system with and without the SA module under AWGN and Rayleigh channels. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: mAP versus SNR for proposed system with and without the KG under AWGN and Rayleigh channels. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: mAP versus SNR for proposed system with and without the multi-scle codec (MS) under AWGN and Rayleigh channels. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Confusion matrices of the proposed system across all fifteen categories. The label of the horizontal axis and vertical axis represents the ground truth [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: AP across all fifteen categories under AWGN channel. (R [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Visual comparison between the proposed system and A+F (AWGN channel, R = 1/6, SNR = -3dB). The first column presents the ground truth, the [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

  1. [9]

    Opportunistic spectrum access in cognitive radio networks: Global optimization using local interaction games,

    Y . Xu, J. Wang, Q. Wu, A. Anpalagan, and Y .-D. Yao, “Opportunistic spectrum access in cognitive radio networks: Global optimization using local interaction games,” IEEE J. Sel. Topics Signal Process. , vol. 6, no. 2, pp. 180–194, 2012

    work page 2012

  2. [35]

    Real-time scene text detection with differentiable binarization and adaptive scale fusion,

    M. Liao, Z. Zou, Z. Wan, C. Yao, and X. Bai, “Real-time scene text detection with differentiable binarization and adaptive scale fusion,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 1, pp. 919–931, 2023

    work page 2023

  3. [37]

    A survey on knowledge graph-based recommender systems,

    Q. Guo, F. Zhuang, C. Qin, H. Zhu, X. Xie, H. Xiong, and Q. He, “A survey on knowledge graph-based recommender systems,” IEEE Trans. Knowledge Data Eng. , vol. 34, no. 8, pp. 3549–3568, 2022

    work page 2022

  4. [1]

    Knowledge Graph Driven UAV Cognitive Semantic Communication Systems for Efficient Object Detection

    X. Song, L. Yuan, Z. Qu, F. Zhou, Q. Wu, T. Q. Quek, and R. Q. Hu, “Knowledge graph driven UA V cognitive semantic communication systems for efficient object detection,” arXiv preprint arXiv:2401.13995, 2024

    work page Pith review arXiv 2024

  5. [2]

    Vehicle detection from UA V imagery with deep learning: A review,

    A. Bouguettaya, H. Zarzour, A. Kechida, and A. M. Taberkit, “Vehicle detection from UA V imagery with deep learning: A review,”IEEE Trans. Neural Netw. Learn. Syst. , vol. 33, no. 11, pp. 6047–6067, 2022

    work page 2022

  6. [3]

    Cross- modal object detection via UA V,

    A. Li, S. Ni, Y . Chen, J. Chen, X. Wei, L. Zhou, and M. Guizani, “Cross- modal object detection via UA V,” IEEE Trans. Veh. Technol. , vol. 72, no. 8, pp. 10 894–10 905, 2023

    work page 2023

  7. [4]

    From external interaction to internal inference: An intelligent learning framework for spectrum shar- ing and UA V trajectory optimization,

    R. Ding, F. Zhou, Q. Wu, and D. W. K. Ng, “From external interaction to internal inference: An intelligent learning framework for spectrum shar- ing and UA V trajectory optimization,” IEEE Trans. Wireless Commun. , pp. 1–1, 2024

    work page 2024

  8. [5]

    Real-time aerial detection and reasoning on embedded-UA Vs in rural environments,

    T. Lai, “Real-time aerial detection and reasoning on embedded-UA Vs in rural environments,” IEEE Trans. Geosci. Remote Sens. , vol. 61, pp. 1–7, 2023

    work page 2023

Show all 39 references
  1. [6]

    Multi-agent deep reinforcement learning for task offloading in UA V-assisted mobile edge computing,

    N. Zhao, Z. Ye, Y . Pei, Y .-C. Liang, and D. Niyato, “Multi-agent deep reinforcement learning for task offloading in UA V-assisted mobile edge computing,” IEEE Trans. Wireless Commun. , vol. 21, no. 9, pp. 6949– 6960, 2022

    2022

  2. [7]

    Joint communication and computation de- sign for UA V-enabled aerial computing,

    J. Ji, K. Zhu, and D. Niyato, “Joint communication and computation de- sign for UA V-enabled aerial computing,”IEEE Commun. Mag., vol. 59, no. 11, pp. 73–79, 2021

    2021

  3. [8]

    UA V- aided ultra-reliable low-latency computation offloading in future IoT networks,

    E. E. Haber, H. A. Alameddine, C. Assi, and S. Sharafeddine, “UA V- aided ultra-reliable low-latency computation offloading in future IoT networks,” IEEE Trans. Commun., vol. 69, no. 10, pp. 6838–6851, 2021

    2021

  4. [10]

    Semantic communications: Principles and challenges,

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic communications: Principles and challenges,” arXiv preprint arXiv:2201.01389 , 2021

    2021 arXiv

  5. [11]

    Bandwidth-agile image transmission with deep joint source-channel coding,

    D. B. Kurka and D. G ¨und¨uz, “Bandwidth-agile image transmission with deep joint source-channel coding,” IEEE Trans. Wireless Commun., vol. 20, no. 12, pp. 8081–8095, 2021

    2021

  6. [12]

    Task-oriented image transmission for scene classification in unmanned aerial systems,

    X. Kang, B. Song, J. Guo, Z. Qin, and F. R. Yu, “Task-oriented image transmission for scene classification in unmanned aerial systems,” IEEE Trans. Commun., vol. 70, no. 8, pp. 5181–5192, 2022

    2022

  7. [13]

    Deep learning-empowered semantic communication systems with a shared knowledge base,

    P. Yi, Y . Cao, X. Kang, and Y .-C. Liang, “Deep learning-empowered semantic communication systems with a shared knowledge base,” IEEE Trans. Wireless Commun., pp. 1–1, 2023

    2023

  8. [14]

    Generative joint source-channel coding for semantic image transmission,

    E. Erdemir, T.-Y . Tung, P. L. Dragotti, and D. G¨und¨uz, “Generative joint source-channel coding for semantic image transmission,” IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2645–2657, 2023

    2023

  9. [15]

    Task-oriented semantic communication over rate splitting enabled wireless control systems for URLLC services,

    C. Zeng, J.-B. Wang, M. Xiao, C. Ding, Y . Chen, H. Yu, and J. Wang, “Task-oriented semantic communication over rate splitting enabled wireless control systems for URLLC services,” IEEE Trans. Commun. , vol. 72, no. 2, pp. 722–739, 2024

    2024

  10. [16]

    Cognitive semantic communication systems driven by knowl- edge graph: Principle, implementation, and performance evaluation,

    F. Zhou, Y . Li, M. Xu, L. Yuan, Q. Wu, R. Q. Hu, and N. Al- Dhahir, “Cognitive semantic communication systems driven by knowl- edge graph: Principle, implementation, and performance evaluation,” IEEE Trans. Commun. , vol. 72, no. 1, pp. 193–208, 2024. 15 Fig. 13. Visual compa...

    2024

  11. [17]

    Knowledge enhanced semantic communication receiver,

    B. Wang, R. Li, J. Zhu, Z. Zhao, and H. Zhang, “Knowledge enhanced semantic communication receiver,” IEEE Commun. Lett. , vol. 27, no. 7, pp. 1794–1798, 2023

    2023

  12. [18]

    Wireless image transmission using deep source channel coding with attention modules,

    J. Xu, B. Ai, W. Chen, A. Yang, P. Sun, and M. Rodrigues, “Wireless image transmission using deep source channel coding with attention modules,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 4, pp. 2315–2328, 2021

    2021

  13. [19]

    Real-time, cloud-based object detection for unmanned aerial vehicles,

    J. Lee, J. Wang, D. Crandall, S. ˇSabanovi´c, and G. Fox, “Real-time, cloud-based object detection for unmanned aerial vehicles,” in Proc. IEEE Int. Conf. Robot. Comput. , 2017, pp. 36–43

    2017

  14. [20]

    Offloading deep learning powered vision tasks from UA V to 5G edge server with denoising,

    S. Ozer, H. E. Ilhan, M. A. Ozkanoglu, and H. A. Cirpan, “Offloading deep learning powered vision tasks from UA V to 5G edge server with denoising,” IEEE Trans. Veh. Technol. , vol. 72, no. 6, pp. 8035–8048, 2023

    2023

  15. [21]

    Online object detection task offloading in UA V ad hoc networks,

    S. Wang, Z. Du, Y . Du, and J. Chen, “Online object detection task offloading in UA V ad hoc networks,” inProc. IEEE Int. Conf. Unmann. Syst., 2022, pp. 758–763

    2022

  16. [22]

    Edge- cloud collaborative UA V object detection: Edge-embedded lightweight algorithm design and task offloading using fuzzy neural network,

    Y . Yuan, S. Gao, Z. Zhang, W. Wang, Z. Xu, and Z. Liu, “Edge- cloud collaborative UA V object detection: Edge-embedded lightweight algorithm design and task offloading using fuzzy neural network,” IEEE Trans. Cloud Comput. , vol. 12, no. 1, pp. 306–318, 2024

    2024

  17. [23]

    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, 2022

    2022

  18. [24]

    Deep joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. Burth 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, 2019

    2019

  19. [25]

    A robust semantic text communication system,

    X. Peng, Z. Qin, X. Tao, J. Lu, and L. Hanzo, “A robust semantic text communication system,” IEEE Trans. Wireless Commun., pp. 1–1, 2024

    2024

  20. [26]

    Semantic-preserved communication system for highly efficient speech transmission,

    T. Han, Q. Yang, Z. Shi, S. He, and Z. Zhang, “Semantic-preserved communication system for highly efficient speech transmission,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 245–259, 2023

    2023

  21. [27]

    Task-oriented explainable semantic communications,

    S. Ma, W. Qiao, Y . Wu, H. Li, G. Shi, D. Gao, Y . Shi, S. Li, and N. Al-Dhahir, “Task-oriented explainable semantic communications,” IEEE Trans. Wireless Commun. , vol. 22, no. 12, pp. 9248–9262, 2023

    2023

  22. [28]

    Beyond transmitting bits: Context, seman- tics, and task-oriented communications,

    D. G ¨und¨uz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, seman- tics, and task-oriented communications,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 5–41, 2022

    2022

  23. [29]

    Task-oriented multi-user semantic communications,

    H. Xie, Z. Qin, X. Tao, and K. B. Letaief, “Task-oriented multi-user semantic communications,” IEEE J. Sel. Areas Commun., vol. 40, no. 9, pp. 2584–2597, 2022. 16

    2022

  24. [30]

    Multi-user semantic commu- nications for cooperative object identification,

    Y . Zhang, W. Xu, H. Gao, and F. Wang, “Multi-user semantic commu- nications for cooperative object identification,” in Proc. IEEE Int. Conf. Commun. Workshops, 2022, pp. 157–162

    2022

  25. [31]

    Semantic transfer between different tasks in the semantic communication system,

    Q. Wu, F. Liu, H. Xia, and T. Zhang, “Semantic transfer between different tasks in the semantic communication system,” in Proc. IEEE Wireless Commun. Netw. Conf., 2022, pp. 566–571

    2022

  26. [32]

    Content-aware semantic communica- tion for goal-oriented wireless communications,

    Y . Fu, W. Cheng, and W. Zhang, “Content-aware semantic communica- tion for goal-oriented wireless communications,” in Proc. IEEE Conf. Comp. Commun. Workshops, 2023, pp. 1–6

    2023

  27. [33]

    Semantic communications for future internet: Fundamentals, applications, and challenges,

    W. Yang, H. Du, Z. Q. Liew, W. Y . B. Lim, Z. Xiong, D. Niyato, X. Chi, X. Shen, and C. Miao, “Semantic communications for future internet: Fundamentals, applications, and challenges,” IEEE Commun. Surveys Tuts., vol. 25, no. 1, pp. 213–250, 2023

    2023

  28. [34]

    6g networks: Beyond shannon to- wards semantic and goal-oriented communications,

    E. Calvanese and S. Barbarossa, “6g networks: Beyond shannon to- wards semantic and goal-oriented communications,” arXiv preprint arXiv:2011.14844, 2021

    2011 arXiv

  29. [36]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proc. IEEE Conf. Comp. Vis. Patt. Recogn. , 2018, pp. 7132–7141

    2018

  30. [38]

    Relational graph attention network for aspect-based sentiment analysis,

    K. Wang, W. Shen, Y . Yang, X. Quan, and R. Wang, “Relational graph attention network for aspect-based sentiment analysis,” arXiv preprint arXiv:2004.12362, 2020

    2004 arXiv

  31. [39]

    Detection and tracking meet drones challenge,

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, and H. Ling, “Detection and tracking meet drones challenge,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 11, pp. 7380–7399, 2022. Xi Song received the B.Eng. degree in software engineering from Shandong Jianzhu Universit...

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.