Denoising Contrastive Regularization (DCR)
- DCR is a contrastive-learning based method that makes denoised outputs approach clean images and separate from noisy inputs using feature-space triplet relations.
- It employs a task-specific embedding network (Wnet) with a ratio-based loss to emphasize high-frequency textures and noise reduction in various imaging modalities.
- Empirical results show that DCR improves restoration metrics (PSNR, SSIM, LPIPS) while reducing artifacts and improving edge sharpness in low-light and CT denoising.
Denoising Contrastive Regularization (DCR) denotes a family of contrastive-learning–based regularization strategies in which a denoised or restored output is encouraged to become closer to a clean target and farther from a noisy or degraded input in a learned feature space. In the clearest explicit formulation, introduced for low-light raw image and video denoising, the denoised image is the anchor, the clean image is the positive, and the noisy image is the negative, with the objective defined over features extracted by a task-related embedding network called Wnet (Cui et al., 2023). Closely related constructions appear in total-body low-dose CT denoising, blind denoising of industrial sensor streams, noisy-label learning, diffusion-based robustness, and diffusion-assisted representation learning, although the exact loss forms, feature spaces, and optimization roles differ across these settings (Fu et al., 2021, Langarica et al., 2020, Yi et al., 2022, Lei et al., 22 Jan 2025, Han et al., 5 Mar 2026).
1. Definition, scope, and terminology
In low-light raw denoising, DCR is defined as a regularizer that, in feature space, makes the denoised image closer to the clean image and far away from the noisy image. The same paper states that conventional denoising networks trained with simple pixel-wise losses tend to either under-denoise or over-smooth, especially in ultra-low-light conditions, and positions DCR as a feature-space correction to that failure mode (Cui et al., 2023).
A conceptually similar description appears in total-body low-dose CT denoising, where a contrastive regularization mechanism is introduced so that restored CT images are pulled closer to Normal-dose CT samples and pushed far away from Low-dose CT samples in the latent space (Fu et al., 2021). This establishes the core semantic pattern of DCR: anchor–positive–negative relations are defined by restoration status rather than by augmentation identity or class membership.
The term is not used uniformly across the literature. In semi-supervised learning and noisy-label learning, the relevant papers use “contrastive regularization” or “contrastive regularization on representations” rather than “DCR,” even though the supplied analyses explicitly interpret them as denoising pseudo-labels or denoising representations (Lee et al., 2022, Yi et al., 2022). By contrast, “DCR” is also used for unrelated names such as “Dual Contrastive Regularization” in unpaired image-to-image translation and “Default Completion Repulsion” in diffusion sampling for rare compositional generation (2304.11319, Kang et al., 7 May 2026). This suggests that DCR is best treated as a contextual term whose meaning depends on the paper rather than as a universally standardized objective.
2. Canonical formulation in low-light raw denoising
The canonical denoising formulation appears in “Contrastive Learning for Low-light Raw Denoising” (Cui et al., 2023). The denoising network takes a sequence of five noisy RAW frames and outputs one denoised RAW image. DCR is then applied in a feature space produced by Wnet. The anchor is the denoised image , the positive is the clean image , and the negative is the noisy image .
The paper describes DCR as a multi-layer feature-space regularizer. For each Wnet layer, it defines a similarity between denoised and clean features and a similarity between denoised and noisy features, then uses a ratio-based objective so that the former is encouraged to be high and the latter low. The authors explicitly note that this is not an InfoNCE softmax-like loss. Instead, it is described as a triplet-style contrastive loss expressed through a ratio of pairwise similarities rather than a margin or a temperature-scaled softmax (Cui et al., 2023).
The similarity function is described as the “mean value of pixel-wise cosine similarity with L1 loss.” This hybrid form is intended to promote both directional alignment in feature space and closeness in magnitude. The final training objective combines three terms,
where is the RAW-space reconstruction term, is computed on the first three RAW channels, and is the DCR term (Cui et al., 2023).
A concise summary of the canonical setup is as follows.
| Component | Role | Description |
|---|---|---|
| Anchor | Denoised output | |
| Positive | Clean reference | |
| Negative | Noisy input | |
| Feature space | Task-related embedding | Wnet |
| Objective type | Ratio-based contrastive regularizer | Not InfoNCE; triplet-style in effect |
This formulation differs from standard self-supervised contrastive learning. There is no memory bank, no large negative pool, and no augmentation-pair semantics. The contrast is restoration-specific: denoised outputs are explicitly regularized to approach clean-image structure while departing from noisy-image structure (Cui et al., 2023).
3. Wnet, feature engineering, and optimization design
A central claim of the low-light DCR paper is that feature space matters as much as loss form. Rather than using a generic semantic encoder such as VGG, the method introduces Wnet, a task-related feature embedding network intended to focus on high-frequency texture and noise components. Wnet uses a Haar wavelet transform front-end, emphasizing the HL, LH, and HH subbands, followed by convolutional layers whose intermediate features are used by DCR (Cui et al., 2023).
Wnet is pre-trained to classify clean versus noisy patches or images using cross-entropy loss. Synthetic noise for that pre-training is generated using a prior noise model from Monakhova et al. Once pre-trained, Wnet is frozen and used only as a feature extractor during denoiser training. The details identify five layers used by DCR and assign progressively larger weights to deeper layers:
0
The supplied explanation states that deeper features therefore contribute more strongly to the contrastive loss (Cui et al., 2023).
The denoising backbones in the study are FastDVDnet and a Modified FastDVDnet in which U-Net blocks are replaced with HRNet for better temporal consistency. The paper states that DCR is independent of the exact backbone. The data regime is explicitly low-light RAW: still images are captured on a moonless clear night in 0.6 millilux, and videos are captured under starlight with no moon present and illuminance below 0.001 lux (Cui et al., 2023).
In the CT setting, the analogous architectural pattern is a teacher–student design. The paper proposes a Teacher-Student Consistency Network consisting of teacher and student networks with identical architecture, uses supervision between intermediate features, and adds a contrastive regularization mechanism so that restored CT images approach NDCT and move away from LDCT in latent space. It also introduces a Dynamic Enhancement Module based on attention and deformable convolution (Fu et al., 2021). Because the supplied record for that paper does not provide the full method section, the precise mathematical form of the CT contrastive term is not available in the supplied text.
4. Empirical behavior and ablation evidence
In the low-light raw study, DCR is evaluated against baseline denoisers and against feature losses based on VGG or Wnet without the DCR ratio formulation. On still-image denoising with FastDVDnet, the reported results are: Baseline, PSNR 30.45, SSIM 0.841, LPIPS 0.076; Baseline + (VGG+L1), PSNR 29.06, SSIM 0.823, LPIPS 0.074; Baseline + (Wnet+L1), PSNR 31.94, SSIM 0.854, LPIPS 0.070; Baseline + DCR, PSNR 31.97, SSIM 0.870, LPIPS 0.065 (Cui et al., 2023).
When FastDVDnet is trained on both paired stills and clean videos, the reported numbers are Baseline, PSNR 31.44, SSIM 0.852, LPIPS 0.0696; Baseline + (VGG+L1), PSNR 30.23, SSIM 0.842, LPIPS 0.0635; Baseline + DCR, PSNR 32.23, SSIM 0.853, LPIPS 0.0630 (Cui et al., 2023). The supplied analysis explicitly interprets this as a roughly 0.8 dB PSNR improvement over baseline for mixed still-image and video training.
For the Modified FastDVDnet evaluated on videos, the strongest comparison in the supplied text is Baseline + DCR at PSNR 30.5, SSIM 0.889, LPIPS 0.060, compared with a retrained baseline at PSNR 27.3, SSIM 0.837, LPIPS 0.091 and a reported state-of-the-art baseline at PSNR 27.7, SSIM 0.931, LPIPS 0.078 (Cui et al., 2023). The qualitative observations in the supplied text describe DCR outputs as having lower residual noise, better edge sharpness, less color bias, clearer objects and silhouettes, less flickering, and fewer artifacts.
The most important ablation logic in that paper is structural rather than purely numerical. The comparison among Baseline, Baseline + (VGG+L1), Baseline + (Wnet+L1), and Baseline + DCR is used to isolate three effects: a generic semantic feature space can be harmful for low-level RAW denoising; a task-related embedding space is better; and the ratio-based contrastive structure adds gains beyond simple feature L1 matching (Cui et al., 2023).
A related empirical pattern appears outside raw denoising. In industrial sensor denoising, the Contrastive Blind Denoising Autoencoder combines a reconstruction loss with an InfoNCE-style latent regularizer and, in the simulated quadruple-tank process, reports RMSE 0.489 cm for CBDAE versus 0.779 cm for the same recurrent autoencoder without regularization and 0.566 cm for an 1-regularized variant at noise power 2 (Langarica et al., 2020). This does not use the exact DCR name, but it provides direct denoising evidence for contrastive regularization as a restoration prior.
5. Generalizations across domains and objective families
Later work broadens the DCR idea beyond paired image denoising. In randomized smoothing and certified robustness, the Robust Representation Consistency Model reformulates denoising along diffusion trajectories as a discriminative task in latent space. Positive pairs are temporally adjacent points on the same probability-flow trajectory, negatives are points from different trajectories, and the model achieves certified-accuracy gains over diffusion-based methods on ImageNet by 5.3% on average, with up to 11.6% at larger radii, while reducing inference costs by 853 on average (Lei et al., 22 Jan 2025). Here the “denoising” object is representation consistency across noise levels rather than image restoration itself.
In “Denoising as Adaptation,” diffusion is used as a domain-adaptation regularizer for image restoration. The method defines a diffusion loss on noisy versions of clean images, conditions the diffusion model on restored synthetic and real outputs, and adds residual-swapping contrastive learning together with channel shuffling. On SIDD denoising, the reported progression is Vanilla U-Net at PSNR 26.58, full-noise-range diffusion without channel shuffling or residual swapping at 32.07, with channel shuffling at 32.91, and the full method at 34.71 (Liao et al., 2024). In this setting, denoising and contrastive regularization jointly align restored synthetic and real outputs with a clean target distribution.
In representation learning, Diffusion Contrastive Reconstruction uses a frozen diffusion model as a teacher and defines an InfoNCE-style loss directly on predicted diffusion noise rather than on original inputs. The paper argues that the resulting DCR loss jointly optimizes discriminative ability and detail perceptual ability, and reports MMVP-VLM gains such as 33.3% for OpenAI ViT-L@224 versus 19.2% for the original encoder (Han et al., 5 Mar 2026). This is no longer denoising in the narrow sense of image cleanup, but it preserves the same principle: a denoising process supplies the geometry in which contrastive regularization is imposed.
The same pattern appears, with different terminology, in learning with noisy labels and semi-supervised learning. “Contrastive Regularization for Semi-Supervised Learning” constructs pseudo-label clusters so that confident anchors pull unconfident positives toward the cluster center and push away negatives, thereby propagating reliable information into more unlabeled samples (Lee et al., 2022). “On Learning Contrastive Representations for Learning with Noisy Labels” proposes Contrastive Regularization on Representations and theoretically argues that the learned representations keep information related to true labels and discard information related to corrupted labels (Yi et al., 2022). These papers do not define DCR as an image-denoising loss, but the supplied analyses explicitly interpret their regularizers as denoising pseudo-labels or denoising representations.
6. Limitations, ambiguities, and open directions
The low-light raw DCR paper identifies several practical constraints. It adds computational overhead because Wnet must process noisy, denoised, and clean images and because multi-layer feature similarities must be computed. It is also sensitive to Wnet quality: if Wnet is poorly trained or not truly noise-aware, DCR may push the denoiser toward suboptimal feature geometry. The method is described as task-specific and tailored for low-light RAW denoising, so applying it to other restoration tasks may require redesign of the embedding network and pre-training targets (Cui et al., 2023).
Across later formulations, the limitations shift with the domain. In diffusion-based robustness, semantic mismatch at high noise levels limits how far consistency can be preserved, and certification cost still scales with the number of noise samples even though per-sample inference is cheaper than diffusion purification (Lei et al., 22 Jan 2025). In diffusion-based domain adaptation, the approach incurs added training cost and memory consumption and depends on a jointly trained diffusion proxy (Liao et al., 2024). In diffusion contrastive reconstruction for CLIP enhancement, the method depends on a pretrained diffusion backbone, and the theoretical guarantees rely on approximations such as a bi-Lipschitz assumption (Han et al., 5 Mar 2026).
A further source of ambiguity is terminological. “Dual Contrastive Regularization” in image-to-image translation is a semantic-plus-style global regularizer, not a denoising objective, even though the supplied discussion notes that the same structure could plausibly be adapted to denoising by using clean images as positives and noisy images as negatives (2304.11319). “DCR: Counterfactual Attractor Guidance for Rare Compositional Generation” uses the acronym for “Default Completion Repulsion,” a projection-based mechanism that removes guidance components aligned with a counterfactual drift direction during diffusion sampling (Kang et al., 7 May 2026). Consequently, the acronym alone does not identify a single method.
Taken together, the literature supports a narrower and a broader reading. In the narrow sense, Denoising Contrastive Regularization is the low-light raw denoising regularizer that uses Wnet to pull denoised outputs toward clean references and away from noisy inputs (Cui et al., 2023). In the broader sense, the term names a design pattern in which denoising structure, restoration state, or noise trajectories define contrastive relations, and those relations regularize a model toward cleaner, more stable, or more semantically faithful representations. This broader interpretation is not a formal standard, but it is the most consistent way to understand the family resemblance across the cited work (Fu et al., 2021, Langarica et al., 2020, Lei et al., 22 Jan 2025, Han et al., 5 Mar 2026).