Semantic-Enhanced Efficient Denoiser (SEED)

• The paper introduces a receiver-side latent denoising method that leverages iterative residual prediction to improve semantic fidelity in image transmission.
• It combines semantic enhancement—via cosine similarity guided updates—with channel-aware conditioning to adaptively reduce noise under low-SNR conditions.
• Empirical results demonstrate significant PSNR and MS-SSIM gains over traditional methods, validating SEED's efficiency and restoration quality.

Searching arXiv for the specified papers to ground the article and verify metadata. Semantic-Enhanced Efficient Denoiser (SEED) can be understood as a class of denoisers that combine semantic-aware recovery with explicit efficiency constraints. In the current literature, the closest concrete instantiation is receiver-side latent denoising for image semantic communication, where denoising is performed on the received semantic latent before semantic decoding, using iterative residual refinement and a similarity-guided adaptive stopping rule. Related restoration work broadens the concept by showing that semantic information can also be injected through training-only semantic statistics matching or through segmentation-conditioned feature modulation, while preserving deployment practicality (Xu et al., 11 Feb 2025, Mei et al., 2022, Xu et al., 2021).

1. Conceptual basis and problem setting

SEED is motivated by a shift from classical bit-level communication and restoration objectives toward semantic fidelity. In image semantic communication, the transmitted object is a learned semantic latent produced by a joint source-channel coding (JSCC) encoder rather than a conventional compressed bitstream with separate channel coding. For an RGB image $\mathbf{x}\in\mathbb{R}^k$, the transmitter maps the image to a latent

$$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$

Two real entries are grouped into one complex symbol to form the channel input $\mathbf{y}_c\in\mathbb{C}^n$, which is transmitted over an AWGN channel,

$$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$

After reshaping back to real form, the receiver obtains a noisy latent $\mathbf{z}_1\in\mathbb{R}^{2n}$ (Xu et al., 11 Feb 2025).

Within this setting, semantic communication is explicitly contrasted with classical information-theoretic systems: the objective is to minimize semantic error, not BER. For image transmission, this means preserving perceptual quality and content-consistent reconstruction rather than only minimizing pixelwise distortion. The central SEED premise is therefore that denoising should act where semantic structure is already compactly encoded. Instead of forcing the semantic decoder to handle both semantic reconstruction and channel corruption simultaneously, a denoiser is inserted as a preprocessing module at the receiver so that the decoder receives a latent closer to the original semantic representation (Xu et al., 11 Feb 2025).

This framing also clarifies the distinction between semantic fidelity and pixel fidelity. Pixel fidelity corresponds to exact or MSE-based image recovery, whereas semantic fidelity refers to preserving the meaningful content encoded in the latent. A SEED-style design therefore favors latent-space denoising over image-space denoising when the task is semantic communication, because the latent is both lower-dimensional and more directly aligned with the transmitted semantics (Xu et al., 11 Feb 2025).

2. Receiver-side latent denoising architecture

The canonical SEED-style pipeline in semantic communication contains five stages: semantic encoding, channel mapping and AWGN transmission, noisy latent reception, latent denoising, and semantic decoding. The receiver-side denoiser is inserted between the physical channel and the semantic decoder: $$\hat{\mathbf{y}}=g_d(\mathbf{z}_1,\mathrm{SNR};\boldsymbol{\psi}), \qquad \hat{\mathbf{x}}=f_d(\hat{\mathbf{y}};\boldsymbol{\theta}).$$ This placement is operationally important: the denoiser acts on the latent representation $\mathbf{z}_1$, not on transmitted symbols after modulation in complex form, and not on the reconstructed image (Xu et al., 11 Feb 2025).

The underlying JSCC codec is based on a Swin Transformer architecture. The encoder has $M+1$ stages, consisting of one patch embedding stage followed by hierarchical transformer stages with patch merging; a convolution head adjusts bandwidth by changing the number of output channels. The decoder mirrors this design and replaces patch merging with reverse merging for upsampling. The reported configuration is

$M=3,\qquad [N_1,N_2,N_3,N_4]=[2,2,6,2].$

The channel model is strictly AWGN; no fading or channel mismatch experiments are reported (Xu et al., 11 Feb 2025).

The denoising module itself comprises two subnetworks: a residual predictor $g_r(\cdot,\cdot;\boldsymbol{\psi}_r)$ and a similarity predictor $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$0, with full parameter set $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$1. At iteration $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$2, the current latent is $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$3 and the current predicted similarity score is $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$4. The update is

$$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$5

The predicted residual latent at step $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$6 is

$$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$7

The paper states that it adopts “a single learnable residual mapping network to iteratively predict and remove the channel noise,” which strongly implies parameter sharing across iterations for the residual mapping. The analogous reuse of the similarity predictor is suggested by the iterative formulation rather than explicitly confirmed. Architecturally, the residual predictor uses a U-Net, the basic CGR block combines convolution, group normalization, and ReLU, and the similarity score is integrated into the upsampling process of the U-Net. The similarity predictor uses convolution layers with $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$8 kernels, while the other neural networks use $$\mathbf{y}=f_e(\mathbf{x};\boldsymbol{\varphi}), \qquad \mathbf{y}\in\mathbb{R}^{2n}.$$9 kernels (Xu et al., 11 Feb 2025).

3. Modes of semantic enhancement

The primary semantic mechanism in latent-space SEED is latent similarity guidance. The denoiser is semantic-aware first because it operates on the semantic latent itself, and second because its denoising process is conditioned and supervised by a latent similarity score defined through cosine similarity between the clean transmitted latent $\mathbf{y}_c\in\mathbb{C}^n$0 and the denoised latent $\mathbf{y}_c\in\mathbb{C}^n$1: $\mathbf{y}_c\in\mathbb{C}^n$2 The corresponding SS loss is added to latent-space MSE during denoiser training. The stated rationale is that MSE alone promotes “pixel-wise average of the plausible latents,” causing overly smooth reconstructions, whereas SS acts as a semantic consistency regularizer and helps recover more image details under low SNR (Xu et al., 11 Feb 2025).

A broader SEED interpretation is supported by restoration literature outside semantic communication. "Deep Semantic Statistics Matching (D2SM) Denoising Network" treats semantic enhancement as training-time regularization rather than an inference-time module. It uses a pretrained classification network as a frozen semantic feature extractor, employs the penultimate layer of VGG19 pretrained on ImageNet for Cityscapes denoising, and minimizes a KL divergence over conditional neighbor-selection probabilities in semantic feature space. For complex scenes, it introduces patch-wise internal probability estimation through sliding-window patches. Its central result is that semantic feature distribution matching can improve denoising while introducing no additional learnable parameters into the deployed denoiser and no inference-time semantic overhead (Mei et al., 2022).

A second complementary line is semantics-conditioned feature modulation. "Synergy Between Semantic Segmentation and Image Denoising via Alternate Boosting" uses semantic probability maps as internal conditioning signals inside the denoiser via Spatial Feature Transform (SFT) layers: $\mathbf{y}_c\in\mathbb{C}^n$3 Here, $\mathbf{y}_c\in\mathbb{C}^n$4 and $\mathbf{y}_c\in\mathbb{C}^n$5 are predicted from the semantic map, so the denoiser becomes class-aware and region-adaptive. The method alternates segmentation and denoising in multiple Segmentation and Denoising Blocks (SDBs), showing that semantic predictions can regularize denoising and that denoised outputs can in turn improve later segmentation stages (Xu et al., 2021).

Taken together, these results indicate that “semantic enhancement” in a SEED context need not refer to a single mechanism. It may denote latent self-consistency in semantic communication, semantic distribution matching during training, or semantic-map conditioning inside the denoiser. A plausible implication is that SEED is better understood as a design family than as one fixed architecture.

4. Efficiency mechanisms and computational profile

Efficiency is a defining component of SEED rather than a secondary implementation detail. In latent semantic communication, the efficiency gain comes first from replacing diffusion-style denoising with direct residual iterative correction. Instead of a lengthy reverse sampling schedule, the denoiser applies a small number of residual updates,

$\mathbf{y}_c\in\mathbb{C}^n$6

which is described as faster and more stable than diffusion-based latent denoisers, especially under low-SNR conditions (Xu et al., 11 Feb 2025).

The second efficiency mechanism is channel-aware conditioning. The denoiser initializes its first similarity score from channel SNR, and the paper states that SNR is used “to estimate and predict the latent similarity score (SS) for conditional denoising.” After each iteration, the similarity predictor estimates the next score, and the predicted SS sequence determines whether additional denoising steps are worthwhile. The stopping rule is qualitative but explicit: continue while the predicted sequence is monotonic non-decreasing, and stop when the SS decreases or when $\mathbf{y}_c\in\mathbb{C}^n$7 is reached. This converts semantic consistency into a self-monitoring signal for adaptive computation (Xu et al., 11 Feb 2025).

The latency evidence is direct. Receiver-side processing times are reported as about $\mathbf{y}_c\in\mathbb{C}^n$8–$\mathbf{y}_c\in\mathbb{C}^n$9 ms for JSCC-T, about $$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$0–$$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$1 ms for the proposed denoiser, and about $$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$2–$$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$3 ms for CDDM. The denoiser therefore adds moderate overhead relative to plain JSCC, but remains orders of magnitude faster than diffusion denoising while improving reconstruction quality. The paper also notes that adaptive inference reduces denoising steps under favorable conditions, making latency savings more noticeable in practical large-volume transmission (Xu et al., 11 Feb 2025).

A related but analogical efficiency vocabulary appears in "SemantiCache: Efficient KV Cache Compression via Semantic Chunking and Clustered Merging." That work is not a denoising method, but it is relevant to SEED at the level of design philosophy. It preserves semantic integrity by delimiter-based chunking, applies Greedy Seed-Based Clustering (GSC) within each chunk, merges clusters into semantic cores by mean pooling, and compensates for aggregation by proportional attention. The reported decoding acceleration reaches up to $$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$4. Because this is KV-cache compression rather than restoration, its relevance to SEED is conceptual rather than direct, but it reinforces a semantic-first view of efficiency and provides a concrete “seed-based” grouping mechanism that can be transferred analogically to denoising or latent aggregation (Wu et al., 15 Mar 2026).

5. Optimization protocol and empirical evidence

The latent denoising formulation most closely associated with SEED is trained in three phases rather than through joint end-to-end optimization from scratch. First, the semantic encoder and decoder are trained over the physical channel using end-to-end image MSE. Second, the pretrained encoder is frozen and the denoiser is inserted. The residual predictor is trained with latent-space MSE plus SS loss, with loss weight

$$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$5

The similarity predictor is trained to estimate the true cosine similarity using MSE. Third, all modules are concatenated, the encoder remains frozen, and only the JSCC decoder is finetuned using end-to-end image MSE again. The denoiser is therefore trained separately after pretraining the codec, followed by decoder finetuning (Xu et al., 11 Feb 2025).

The reported experimental setup uses DIV2K for training, with 800 images at 2K resolution, and Kodak24 for testing, with 24 RGB images at resolution $$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$6. The bandwidth ratio is

$$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$7

and the main reported reconstruction results use

$$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$8

The channel model is AWGN. The first training phase uses SNR $$\mathbf{z}_c=\mathbf{y}_c+\mathbf{n}_c,\qquad \mathbf{n}_c\sim \mathcal{CN}(\mathbf{0},\sigma_n^2\mathbf{I}_n).$$9 dB; the second and third training phases use $\mathbf{z}_1\in\mathbb{R}^{2n}$0 dB; evaluation includes results down to $\mathbf{z}_1\in\mathbb{R}^{2n}$1 dB. Baselines include SSCC with BPG + LDPC + 4QAM, the transformer-based JSCC system without the latent denoiser (JSCC-T), and CDDM. Evaluation metrics are PSNR, MS-SSIM, and receiver-side latency (Xu et al., 11 Feb 2025).

The main empirical result is that residual iterative latent denoising improves both reconstruction quality and efficiency in low-SNR semantic communication. At $\mathbf{z}_1\in\mathbb{R}^{2n}$2 under AWGN, the proposed method consistently outperforms JSCC-T and CDDM on PSNR; at SNR $\mathbf{z}_1\in\mathbb{R}^{2n}$3 dB, the gain is reported as $\mathbf{z}_1\in\mathbb{R}^{2n}$4 dB over JSCC-T and $\mathbf{z}_1\in\mathbb{R}^{2n}$5 dB over CDDM. For MS-SSIM, the method outperforms the others under most SNRs. Ablations show that the empirical initialization of the first similarity score outperforms fixed or random alternatives, that adaptive inference gives near-optimal performance under fluctuating SNR while reducing average denoising steps, and that adding SS loss improves high-frequency detail recovery; at SNR $\mathbf{z}_1\in\mathbb{R}^{2n}$6 dB, the reported example is $\mathbf{z}_1\in\mathbb{R}^{2n}$7 dB with SS loss versus $\mathbf{z}_1\in\mathbb{R}^{2n}$8 dB without it (Xu et al., 11 Feb 2025).

Supporting restoration results indicate that semantic enhancement is not confined to semantic communication. On Cityscapes denoising, D2SM improves PSNR, SSIM, and MIoU across FFDNet, CBDNet, and SADNet, with the “w/. Internal” variant consistently better than “w/o. Internal” on complex scenes. In alternate boosting, SDB$\mathbf{z}_1\in\mathbb{R}^{2n}$9 on Cityscapes at $$\hat{\mathbf{y}}=g_d(\mathbf{z}_1,\mathrm{SNR};\boldsymbol{\psi}), \qquad \hat{\mathbf{x}}=f_d(\hat{\mathbf{y}};\boldsymbol{\theta}).$$0 reports $$\hat{\mathbf{y}}=g_d(\mathbf{z}_1,\mathrm{SNR};\boldsymbol{\psi}), \qquad \hat{\mathbf{x}}=f_d(\hat{\mathbf{y}};\boldsymbol{\theta}).$$1 for PSNR/SSIM and mIoU $$\hat{\mathbf{y}}=g_d(\mathbf{z}_1,\mathrm{SNR};\boldsymbol{\psi}), \qquad \hat{\mathbf{x}}=f_d(\hat{\mathbf{y}};\boldsymbol{\theta}).$$2, improving on both single-stage baselines and plain cascaded denoisers. These results collectively support the SEED premise that denoising quality should be assessed not only by distortion metrics but also by preservation of semantically discriminative structure (Mei et al., 2022, Xu et al., 2021).

6. Scope, limitations, and prospective generalization

Current SEED-style methods remain specialized. The latent semantic communication formulation is evaluated only under AWGN, with no fading, burst noise, or channel mismatch analysis; it is image-only, with no evidence on other modalities or downstream semantic tasks; it provides no explicit theoretical convergence guarantee for iterative residual updates; and its adaptive stopping rule is heuristic, based on a predicted decrease in SS. The initialization of the first similarity score is empirical because the exact closed-form cosine similarity under AWGN is said to be intractable, and step scheduling may be sensitive to the accuracy of the similarity predictor (Xu et al., 11 Feb 2025).

The semantic-regularization and segmentation-coupled lines have different constraints. D2SM adds training overhead through pretrained semantic feature extraction, pairwise similarity computation, and feature queues, yet does not provide explicit runtime, FLOPs, or memory comparisons. It also omits the exact weighted total-loss specification. The alternate boosting architecture demonstrates semantic-denoising synergy, but efficiency is not its primary focus: repeated HRNet and DDFN stages increase inference cost, and the experiments are based mainly on synthetic AWGN and Poisson noise rather than real noisy datasets with semantic labels (Mei et al., 2022, Xu et al., 2021).

The analogical seed-based compression line is also bounded by its own assumptions. Delimiter-based chunking is heuristic, clustering uses keys only, GSC is local and order-dependent, mean pooling may blur fine-grained distinctions, and proportional attention is heuristic rather than exact. For denoising, these mechanisms are therefore suggestive rather than validated transfer results (Wu et al., 15 Mar 2026).

Several generalizations have been explicitly proposed for a more general SEED framework. These include broader semantic supervision through perceptual or foundation-model feature losses such as CLIP or ViT feature similarity, richer conditioning using estimated channel statistics or uncertainty, dynamic computation via confidence-based halting, multi-objective semantic fidelity that includes downstream task accuracy such as classification or segmentation, general channel robustness beyond AWGN, and parameter-efficient iterative refinement with lightweight recurrent memory or state compression. This suggests that SEED is best viewed as an evolving research direction centered on three stable principles: denoise before semantic decoding when semantic latents are available, preserve semantic structure explicitly during optimization or conditioning, and adapt computation to the corruption level rather than applying uniform denoising effort in all cases (Xu et al., 11 Feb 2025).