Digital-Analog Transmission based Emergency Semantic Communications

Jingqing Wang; Liuguo Yin; Wei Zhang; Wenchi Cheng; Yuzhou Fu

REVIEW 4 major objections 8 minor 6 references

Digital-Analog Transmission based Emergency Semantic Communications

T0 review · 4 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A hybrid analog-digital semantic codec with a CRLB-trained encoder and distributed-source-coding parity refinement outperforms classical separated and purely neural joint source-channel coding for emergency image transmission.

desk verdict Hybrid digital-analog semantic transmission is a plausible engineering idea, but the CRLB that anchors the theory is invalid, so the paper needs major revision. read the letter →

arxiv 2501.01616 v1 pith:L5Q7MRAP submitted 2025-01-03 eess.SP

Yuzhou Fu , Wenchi Cheng , Jingqing Wang , Liuguo Yin , Wei Zhang This is my paper · ORCID

classification eess.SP

keywords emergencywirelesscommunicationsemanticdigital-analogtransmissionjointsource-channelcodingdistributedsourceCramér-Raolowerboundimagereconstructionobjectdetection

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 hybrid transmission scheme for emergency wireless image delivery, where the transmitter has limited neural-network capacity and UDP provides no retransmissions. The scheme pairs an analog deep-learning semantic encoder with a digital refinement path: analog symbols carry a coarse semantic estimate, and low-density parity-check digits from distributed source coding correct that estimate at the receiver. To train the analog codec, the paper derives a Cramér-Rao lower bound for the combined semantic-noise and channel-noise distortion and uses that bound, together with reconstruction error, as the optimization objective. In simulations over fading channels, the hybrid scheme achieves better reconstruction fidelity (PSNR, LPIPS, MS-SSIM) and better fire-detection scores (mAP50, mAP50:95) than classical separated source-channel coding and than DL-based joint source-channel coding at all tested SNR values. The claim is that this removes the fidelity ceiling caused by constrained neural-network capacity without requiring retransmission or a large model.

What carries the argument

The load-bearing identity is the Cramér-Rao lower bound of the overall distortion, expressed in Eq. (16) as a sum over transmitted patches of a term involving the expected squared gradient of the semantic encoder mapping and a sum over synonymously mapped patches of a term governed by the fitting coefficient $\rho_i^2$. The paper uses this expression as the optimization objective for training the semantic codec. The second mechanism is the DA-ESemCom architecture: an analog semantic encoder selects $Q$ image patches with the lowest fitting coefficients and maps them to channel symbols, a powerful decoder forms a coarse estimate, and a digital DSC path transmits LDPC parity bits derived from DCT-quantized sources, treating the estimate as side information. A lightweight resource-allocation rule gives digital transmission the power and bandwidth needed for a target error probability and gives the remaining bandwidth to the analog path.

What would settle it

Train the same semantic codec with the true end-to-end mean squared error between source and final reconstruction as the loss and compare it against the proposed CRLB-based loss on the paper's fading-channel setup; if the CRLB-based loss does not improve PSNR or mAP, or if Eq. (11)'s gradient identity fails when complex Rayleigh fading is inserted directly, the claimed role of Eq. (16) as a design objective is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that a small emergency transmitter can exceed the reconstruction ceiling of a purely analog neural semantic codec by adding a low-rate digital correction path. The analog path transmits semantically encoded image patches and yields a coarse estimate at the receiver; the digital path, built on distributed source coding, sends only LDPC parity bits, and the receiver uses the coarse estimate as side information to refine the final reconstruction. The paper further claims that the Cramér-Rao lower bound derived for the overall distortion, Eq. (16), is a valid design objective: training the semantic encoder to minimize that bound along with reconstruction MSE improves adaptivity to both semantic noise and channel noise. In the reported simulations, the proposed DA-ESemCom framework outperforms classical SSCC and two DL-based JSCC baselines in reconstruction quality and object-detection performance, and it even exceeds the PSNR of an ideal noiseless semantic-transmission baseline once the SNR passes about 2 dB.

Load-bearing premise

The derivation's load-bearing premise is that the distortion in Eq. (4), and hence the Cramér-Rao lower bound and the training loss built from it, is computed for the error between the encoder output and the received symbol, with the fading coefficient treated as a real unit-power scalar, rather than for the true end-to-end reconstruction error under complex Rayleigh fading.

Editorial extensions

If this is right

Constrained transmitters can gain most of the benefit of a large neural codec: in the ablation, the proposed framework with the small base model improves MS-SSIM over DL-based JSCC with the large model by roughly 0.3 dB.
Reconstruction fidelity stops saturating at high SNR: the hybrid scheme exceeds the paper's ideal noiseless semantic-transmission baseline in PSNR once the SNR exceeds about 2 dB.
AI-driven fire detection benefits across the whole SNR range: both mAP50 and mAP50:95 scores improve over the baselines and approach the scores of original images at high SNR.
The classical cliff effect of separated source-channel coding can be avoided while retaining digital fidelity, because the analog path provides graceful degradation at low SNR and the DSC path adds precision at high SNR.
The framework works without retransmission, making it compatible with the UDP transport used in emergency wireless networks.

Reading between the lines

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

A testable extension the paper does not report: an ablation of DA-ESemCom trained with plain reconstruction MSE versus the proposed CRLB-based loss would isolate how much of the gain comes from the digital parity path and how much from the new training objective.
The same analog-plus-digital-parity idea could transfer to other task-oriented emergency sensing modalities, such as compressed video or point-cloud frames, wherever a coarse semantic estimate can serve as side information for distributed source coding.
The paper's resource-allocation rule is heuristic; a joint optimization over the number of analog patches $Q$, the DSC coding rate, and the power split would likely improve on the priority rule, but the paper does not claim such optimality.
A useful diagnostic: comparing the CRLB-based loss against the true end-to-end reconstruction error would show whether Eq. (16) mainly regularizes the encoder or directly drives the reported fidelity gains.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 — here's the short version: the DA-ESemCom framework is a sensible extension of the known hybrid digital-analog idea to emergency image transmission, and the simulations suggest it works. But the CRLB derivation in Section II-C, which is one of the paper's two main contributions, does not hold up. The distortion in Eq. (4) is defined as E[(f_E(s_i)-y_i)^2] — the error between the encoder output and the received symbol — not the end-to-end reconstruction error E[(s_i - \hat{s}_i)^2]. The CRLB is a bound on the variance of an unbiased estimator of s_i; it cannot bound that surrogate quantity. The fading coefficient is also mishandled: with h_i ~ CN(0,1), E[h_i^2]=0, so replacing E[(∂h_i f_E/∂s)^2] with E[(∂f_E/∂s)^2] in Eq. (11) is wrong. And Eq. (16) does not follow arithmetically from Eq. (13). Because the training loss in Eq. (18) includes the CRLB-derived term, the 'DL-based JSCC (Proposed loss)' baseline is partly testing the loss that is supposed to be justified by the theory. That is circular.

What is genuinely new: the emergency-communication framing, the synonymous-mapping model for dropped patches, and the lightweight power/bandwidth allocation that prioritizes the digital parity part. The paper does a decent job on the empirical side — PSNR, LPIPS, mAP, visual examples, and an ablation over model sizes. The authors also state the inherent fidelity bottleneck honestly. But no code or data are released, and the 'ideal semantic transmission' upper bound is not specified tightly enough to be a fair comparison.

Proportionate bottom line: the system idea may be sound and the empirical gains plausible, but the theoretical contribution as written is not valid. This is a paper for the semantic-communications and emergency-wireless crowd; they will find the architecture and the resource allocation interesting, but they should not cite the CRLB result.

My recommendation: don't desk-reject — the empirical work deserves scrutiny and the CRLB can be fixed or dropped. But I would send it back for major revision, and before I'd trust the numbers I'd want code/data and a fairer baseline.

Referee Report

4 major / 8 minor

Summary. The paper proposes DA-ESemCom, a digital-analog transmission framework for emergency wireless communications. The analog part is a deep-learning-based semantic codec with a performance-constrained model that incorporates semantic noise and channel noise; the digital part is a distributed source coding (DSC) scheme based on LDPC parity bits that refines the analog reconstruction. The authors derive a Cramér–Rao lower bound (CRLB) for the proposed semantic coding model, use it to define a training loss, and present simulations showing that the full framework outperforms classical SSCC (JPEG+LDPC) and DL-based JSCC baselines in PSNR, LPIPS, MS-SSIM, and object-detection mAP over a Rayleigh fading channel. The paper also includes an ablation study on model size.

Significance. The problem addressed is relevant and timely: emergency transmitters with limited hardware, no ARQ, and low SNR can benefit from robust semantic communication. The hybrid analog-digital architecture—using DSC parity to refine a semantic analog reconstruction—is a sensible and potentially useful engineering idea, and the empirical comparisons cover both fidelity and task-oriented detection metrics. The paper also attempts a theoretical justification of the codec design via CRLB, which, if valid, would be a valuable design principle. However, the theoretical derivation contains load-bearing errors: the distortion being bounded is not the end-to-end reconstruction error, the fading coefficient is treated inconsistently with the stated complex Gaussian model, and the final CRLB expression does not follow algebraically from the preceding equations. Because the CRLB is the basis for the training objective and the claimed performance bottleneck, the central theoretical contribution is not supported. The empirical results remain suggestive, but they are not sufficient to salvage the paper's main claims as written.

major comments (4)

[Section II-C, Eq. (4)–(5)] The distortion D_overall in Eq. (4) is defined as the expected squared error between the encoder output f_E(s_i) and the received symbol y_i, not as the end-to-end reconstruction error E[(s_i - \hat{s}_i)^2]. The CRLB in Eq. (5) is a lower bound on the variance of an unbiased estimator of a parameter (here s_i) from the observation y_i. It cannot be applied to this surrogate quantity, which is not an estimator variance. Consequently, the bound in Eq. (5) and the training objective (18) built on it do not lower-bound the actual image reconstruction distortion.
[Section II-C, Eqs. (1), (6), (9)–(11)] The complex fading coefficient h_i ~ CN(0,1) is mishandled. In Eq. (1), the effective noise is h_i \epsilon_n + n, whose conditional variance given h_i is |h_i|^2 \sigma_\epsilon^2 + \sigma_n^2, not the value \sigma_\epsilon^2/(1+\sigma_\epsilon^2+\sigma_n^2) used in Eq. (6). Moreover, for h_i ~ CN(0,1), E[h_i^2] = 0, so the replacement in Eq. (11) of E[(\partial(h_i f_E)/\partial s)^2] by E[(\partial f_E/\partial s)^2] is invalid. The Fisher information for this complex channel model should involve |h_i|^2 or an expectation over |h_i|^2; the derivation as written effectively treats h_i as a real unit-power scalar. This error propagates into Eqs. (13) and (16).
[Section II-C, Eqs. (13)–(16)] Equation (16) does not follow algebraically from Eq. (13). From Eq. (13), the first term is \sigma_\epsilon^2 / [(1+\sigma_\epsilon^2+\sigma_n^2) \sum_l E[(\partial f_E/\partial s)^2]], whereas Eq. (16) gives a different numerator (\sigma_\epsilon^2(1+\sigma_n^2)+\sigma_n^2(1+\sigma_\epsilon^2)). The second term in Eq. (13) is \sigma_{t,i}^2 \sigma_{s,i}^2 \sigma_\epsilon^2 / L, which with \rho_i = \sigma_{t,i}/\sigma_{s,i} equals \rho_i^2 \sigma_{s,i}^4 \sigma_\epsilon^2 / L, not \sigma_\epsilon^2/(L \rho_i^2) as in Eq. (16). The design insight that \rho_i^2 should be maximized and the subsequent loss (18) are therefore based on an incorrect formula.
[Section II-C, Eq. (8)] The conditional density p(t_i|s_i) is derived incorrectly. From s_i = t_i + \epsilon_n with independent Gaussian t_i and \epsilon_n, the conditional distribution of t_i given s_i is Gaussian with mean (\sigma_t^2/(\sigma_t^2+\sigma_\epsilon^2)) s_i and variance \sigma_t^2 \sigma_\epsilon^2/(\sigma_t^2+\sigma_\epsilon^2), not the expression in Eq. (8). This invalidates the Fisher information computation for the synonymous-mapping term and hence the corresponding CRLB term.

minor comments (8)

[Algorithm 1, line 13] The text says 'Estimate the various of the semantic noise'; 'various' should be 'variance'.
[Algorithm 1 and text] The symbol f_{Ieal}^E should be f_{Ideal}^E; the typo appears in Algorithm 1 and in Section II-D.
[Figures 3–6] The figure captions contain the misspelling 'Porposed' for 'Proposed'.
[Eq. (1)] The acronym 'AWGM' should be 'AWGN'.
[References] Reference [23] is Slepian and Wolf's noiseless coding of correlated sources, which is not the standard citation for the Cramér–Rao lower bound; a proper CRLB reference is needed. Reference [25] on texture stationarity is not the usual basis for the sample-variance formula in Eq. (7).
[Section V-C] The baseline description does not specify the JPEG quality factor, the LDPC code rate, or the symbol-to-bandwidth mapping for the classical SSCC scheme, which are needed to reproduce the comparison.
[Section V-C] The 'ideal semantic transmission' baseline is vaguely defined as 'full, noiseless, and high-dimensional semantic feature vectors with full-resolution constellation'; a quantitative definition of bandwidth, power, and constellation resolution is needed to interpret why the proposed DA-ESemCom can surpass it.
[Section III and Eq. (3)] The notation \hat{S} is used in Eq. (3) for the decoder output, but the final output of the DA-ESemCom framework is also denoted S in Section IV; please disambiguate these.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central DA-ESemCom comparison is externally benchmarked, and the CRLB-derived loss is a design objective rather than a fitted prediction.

full rationale

After walking the derivation chain, I find no step in which a 'prediction' or 'derived bound' is equivalent by construction to its inputs. The paper's headline result—DA-ESemCom outperforms classical SSCC and DL-based JSCC—is established by simulations against independent baselines (JPEG+LDPC SSCC; JSCC trained with MSE loss) on ImageNet-1K and COCO2017, so it would stand or fall on its own. The CRLB expression in Eq. (16) is used as a component of the training loss (18); that makes it a design objective, not a fitted parameter renamed as a prediction. Ablating 'Proposed loss' against 'MSE loss' is an internal comparison and is not the paper's only evidence. The self-citation [26] for the DA framework idea is not load-bearing: the paper also cites independent hybrid digital-analog works [27],[28] and supplies its own Algorithm 2, resource allocation, and simulations. I do note correctness concerns outside this pass: Eq. (4) defines distortion as E[(f_E-y)^2], which is not the variance of an unbiased estimator of s, so the CRLB in Eq. (5)/(16) may not bound the quantity the paper claims; the treatment of the fading coefficient h_i in Eqs. (6) and (11) drops h_i despite h_i ~ CN(0,1); and there is a dangling citation '[A]' after Eq. (30). These are mathematical/completeness issues, not circularity. Accordingly, score 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The paper introduces several modeling constructs and hand-set hyperparameters that the central claims depend on. The semantic noise, synonymous tokens, and ideal encoder are not independently evidenced. The CRLB derivation further relies on an inconsistent fading assumption and an unbiased-decoder assumption that neural networks do not satisfy.

free parameters (4)

Semantic noise variance sigma_eps^2 = estimated per batch via MLE in Algorithm 1, step 13
Used to define likelihoods and CRLB in Eqs. (6)-(16); its value is estimated from the difference between the ideal and constrained encoders during training, so it is data-dependent.
Masking ratio eta = 0.7
Hand-set in Algorithm 1; determines how many patches are sent versus synonymously mapped, strongly affecting the fidelity bottleneck.
Training SNR gamma = 10 dB
Hand-set in Algorithm 1; the codec is trained at this channel SNR, influencing all later SNR comparisons.
Digital error-probability threshold epsilon_th = not specified in the paper
Sets P_d and B_d via Eq. (28) and therefore the analog/digital resource split; the value is not reported in the simulations.

assumptions (4)

domain assumption Image patches are modeled as Gaussian sources (Section II-B).
Real scene images are not Gaussian; the likelihoods and CRLB in Section II-C are built on this model.
domain assumption The semantic decoder mapping is unbiased (Section II-C).
CRLB only bounds unbiased estimators; trained neural decoders are generally biased, so the bound may not apply.
ad hoc to paper Rayleigh fading coefficient satisfies E[h_i^2]=1 despite the stated h_i~CN(0,1).
Eq. (11) keeps E[(d(h f_E)/ds)^2] as if E[h^2]=1, but a circular complex Gaussian has E[h^2]=0; this inconsistency is needed for the CRLB expression to be finite.
domain assumption DSC coding with LDPC parity bits provides the benefit of Slepian-Wolf distributed source coding (Section III).
The cited Slepian-Wolf framework [23] is not actually implemented; the digital part is a conventional channel-code refinement, so the theoretical DSC motivation is assumed by analogy.

invented entities (3)

Semantic noise epsilon_n
purpose: Models errors from finite constellation mapping and synonymous mapping in the semantic encoder (Eqs. (1)-(2)).
No independent measurement is provided; it is a postulated modeling term whose variance is estimated during training.
Synonymous mapping tokens t_i
purpose: Shared representations used to reconstruct dropped image patches (Eq. (2)).
No external validation that such tokens exist or achieve the claimed fit; their distribution is postulated as Gaussian.
Ideal semantic encoder f_ideal_E
purpose: Provides a noiseless reference to estimate semantic noise during training (Algorithm 1, line 8).
A hypothetical construct; no definition or implementation details are given, and its output defines the 'ideal semantic transmission' upper bound.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Digital-Analog Transmission based Emergency Semantic Communications." pith.science (2026). https://pith.science/paper/L5Q7MRAP

@misc{pith2026250101616,
  author       = {Pith},
  title        = {Pith review of: Digital-Analog Transmission based Emergency Semantic Communications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5Q7MRAP}},
  note         = {Machine review of arXiv:2501.01616}
}

read the original abstract

Emergency Wireless Communication (EWC) networks adopt the User Datagram Protocol (UDP) to transmit scene images in real time for quickly assessing the extent of the damage. However, existing UDP-based EWC exhibits suboptimal performance under poor channel conditions since UDP lacks an Automatic Repeat reQuest (ARQ) mechanism. In addition, future EWC systems must not only enhance human decisionmaking during emergency response operations but also support Artificial Intelligence (AI)-driven approaches to improve rescue efficiency. The Deep Learning-based Semantic Communication (DL-based SemCom) emerges as a robust, efficient, and taskoriented transmission scheme, suitable for deployment in UDP based EWC. Due to the constraints in hardware capabilities and transmission resources, the EWC transmitter is unable to integrate sufficiently powerful NN model, thereby failing to achieve ideal performance under EWC scene. For EWC scene, we propose a performance-constrained semantic coding model, which considers the effects of the semantic noise and the channel noise. Then, we derive Cramer-Rao lower bound of the proposed semantic coding model, as guidance for the design of semantic codec to enhance its adaptability to semantic noise as well as channel noise. To further improve the system performance, we propose Digital-Analog transmission based Emergency Semantic Communication (DAESemCom) framework, which integrates the analog DL-based semantic coding and the digital Distributed Source Coding (DSC) schemes to leverage their respective advantages. The simulation results show that the proposed DA-ESemCom framework outperforms the classical Separated Source-Channel Coding (SSCC) and other DL-based Joint Source-Channel Coding (DL-based JSCC) schemes in terms of fidelity and detection performances.

Figures

Figures reproduced from arXiv: 2501.01616 by the authors.

**Figure 2.** The proposed DA-ESemCom framework for enhancing the performance-constrained semantic coding transmission. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Performance of reconstruction results of the proposed DA-ESemCom framework and the other baselines over fading [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Performance on object detection of the proposed DA-ESemCom framework and the other baselines over fading channel. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Visual examples of object detection results on the reconstructed images of the proposed DA-ESemCom framework and [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Ablation study results in terms of the PSNR and the mAP50 performances over the fading channel. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 6 canonical work pages

[1]

Classical SSCC: The classical SSCC is based on Shan- non’s reliable transmission framework, where widely used Joint Photographic Experts Group (JPEG) [38] and LDPC are adopted as the source coding and the channel coding, respectively

work page
[2]

Ideal semantic transmission: The ideal semantic trans- mission scheme, serving as the performance upper bound of the proposed semantic coding scheme, is with unlimited channel resources, where full, noiseless, and high-dimensional semantic feature vectors can be transmitted to the receiver with full-resolution constellation

work page
[3]

Enabling emergency communication through a cognitive radio vehicular network,

DL-based JSCC: The DL-based JSCC schemes adopt the proposed semantic codec without using the DSC part, where DL-based JSCC (Proposed loss) and DL-based JSCC (MSE loss) schemes are trained with the proposed optimiza- tion objective (18) and only MSE objective (17), respectively. D. Objective Quality for Different Methods Figure 3 shows the performance of r...

work page 2014
[36]

The dataset is widely used for various AI-driven vision tasks, such as classiﬁcation, detection, and segmentation

contains 1000 different categories in the training set and each category contains about 1000 images. The dataset is widely used for various AI-driven vision tasks, such as classiﬁcation, detection, and segmentation. We use this dataset to pre-train our proposed semantic coding, laying the basis for learning the optimal coded representations. COCO2017

work page
[37]

This diversity is convenient to evaluate the performance of the system under different ﬁre scenarios

contains 7,157 images and includes dif- ferent kinds of ﬁres such as building, wildland and residential ﬁre, containing shots captured at day time, dusk or night time. This diversity is convenient to evaluate the performance of the system under different ﬁre scenarios. C. The Adopted Comparison Schemes We compare our proposed DA-ESemCom framework with the...

work page
[2014]

Transmission of JPEG- coded images over wireless channels,

[Online]. Available: https://www.microsoft.com/en-us/research/ publication/microsoft-coco-common-objects-in-context/ [38] V. Weerackody, C. Podilchuk, and A. Estrella, “Transmission of JPEG- coded images over wireless channels,” Bell Labs Technical Journal , vol. 1, no. 2, pp. 111–126, 1996

work page 1996

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Classical SSCC: The classical SSCC is based on Shan- non’s reliable transmission framework, where widely used Joint Photographic Experts Group (JPEG) [38] and LDPC are adopted as the source coding and the channel coding, respectively

work page

[2] [2]

Ideal semantic transmission: The ideal semantic trans- mission scheme, serving as the performance upper bound of the proposed semantic coding scheme, is with unlimited channel resources, where full, noiseless, and high-dimensional semantic feature vectors can be transmitted to the receiver with full-resolution constellation

work page

[3] [3]

Enabling emergency communication through a cognitive radio vehicular network,

DL-based JSCC: The DL-based JSCC schemes adopt the proposed semantic codec without using the DSC part, where DL-based JSCC (Proposed loss) and DL-based JSCC (MSE loss) schemes are trained with the proposed optimiza- tion objective (18) and only MSE objective (17), respectively. D. Objective Quality for Different Methods Figure 3 shows the performance of r...

work page 2014

[4] [36]

The dataset is widely used for various AI-driven vision tasks, such as classiﬁcation, detection, and segmentation

contains 1000 different categories in the training set and each category contains about 1000 images. The dataset is widely used for various AI-driven vision tasks, such as classiﬁcation, detection, and segmentation. We use this dataset to pre-train our proposed semantic coding, laying the basis for learning the optimal coded representations. COCO2017

work page

[5] [37]

This diversity is convenient to evaluate the performance of the system under different ﬁre scenarios

contains 7,157 images and includes dif- ferent kinds of ﬁres such as building, wildland and residential ﬁre, containing shots captured at day time, dusk or night time. This diversity is convenient to evaluate the performance of the system under different ﬁre scenarios. C. The Adopted Comparison Schemes We compare our proposed DA-ESemCom framework with the...

work page

[6] [2014]

Transmission of JPEG- coded images over wireless channels,

[Online]. Available: https://www.microsoft.com/en-us/research/ publication/microsoft-coco-common-objects-in-context/ [38] V. Weerackody, C. Podilchuk, and A. Estrella, “Transmission of JPEG- coded images over wireless channels,” Bell Labs Technical Journal , vol. 1, no. 2, pp. 111–126, 1996

work page 1996