DDR-Net for Joint Deraining & Low-Light Enhancement

• The paper introduces DDR-Net, a dual degradation representation network that disentangles luminance and rain features from low-light rainy images.
• DDR-Net employs a dual-branch encoding with spatially adaptive attention that separately processes bright and dark regions to better capture degradation details.
• Empirical results demonstrate that DDR-Net improves restoration metrics like PSNR and SSIM while mitigating artifacts compared to cascade-based approaches.

DDR-Net denotes the Dual Degradation Representation Network introduced as a core component of $L^{2}$RIRNet for joint deraining and low-light enhancement in the dark. In the paper, the same module is described as the Pairwise Degradation Feature Vector Extraction Network (P-Net). Its function is to extract and independently represent two degradations that co-occur in low-light-rainy images—luminance (low-light) degradation and rain pattern degradation—so that a downstream restoration network can remove rain and brighten dark regions within a single end-to-end model rather than through cascade pipelines that often yield problematic rain patterns or overly blurred and overexposed images (Lin et al., 2023).

1. Problem setting and design objective

Rain in the dark is presented as a difficult restoration regime for autonomous driving, surveillance systems, and night photography. The central claim motivating DDR-Net is that existing low-light enhancement methods and deraining methods each address only part of the degradation, while straightforward cascades such as “deraining followed by low-light enhancement” or the reverse can amplify artifacts rather than resolve them (Lin et al., 2023).

Within $L^{2}$RIRNet, DDR-Net is paired with a Restoration Network. The former is responsible for degradation representation learning; the latter performs image restoration guided by those learned codes. The decomposition of labor is explicit: DDR-Net learns representations for luminance effects in dark areas and rain patterns in light areas, and the Restoration Network uses those representations together with Fourier Detail Guidance to reconstruct a clean image (Lin et al., 2023).

A key premise behind DDR-Net is spatial nonuniformity. The paper states that rain streaks are much more visible in bright regions, whereas low-light degradation is dominant in dark regions. It further grounds this separation in a physical model showing that rain streak visibility and the dominant type of degradation vary with distance from light sources. This motivates an architecture that does not treat degradation as spatially homogeneous (Lin et al., 2023).

2. Architecture of the Dual Degradation Representation Network

DDR-Net/P-Net uses a dual-branch encoding strategy with spatially adaptive attention. Its first operation is to split the input image into two complementary spatial regions using an attention map $MAP$ and its complement $1-MAP$. The attention map is taken from EnlightenGAN, and the two masked images are defined as follows:

$$\begin{aligned} P_{s} &= \text{image} \otimes (1-MAP) \ P_{c} &= \text{image} \otimes MAP \ P_{en} &= \text{Concat}(P_{s}, P_{c}) \end{aligned}$$

Here, $P_s$ corresponds to bright regions, and $P_c$ corresponds to dark regions. The bright branch is intended to focus on areas affected by rain patterns, often near light sources, whereas the dark branch emphasizes low-light-degraded regions (Lin et al., 2023).

After concatenation, $P_{en}$ is processed by an encoder-decoder structure (U-Net-like). The encoder projects the concatenated input into a latent degradation feature space, and the decoder produces a multi-channel latent degradation feature vector:

$$\begin{aligned} \text{Feature} &= \text{P\_Encoder}(P_{en}) \ \text{Latent} &= \text{P\_Decoder}(\text{Feature}) \end{aligned}$$

An additional contrastive learning head is attached to the encoded features. Specifically, the encoded representation is projected through a multilayer perceptron for contrastive loss computation. This arrangement makes DDR-Net not merely an encoder of appearance, but an encoder of degradation-specific latent structure (Lin et al., 2023).

3. Independent representation of luminance and rain

The paper characterizes DDR-Net as a mechanism for disentangling luminance and rain features by exploiting the observation that the two degradations dominate in different spatial regions. Processing bright and dark regions separately with attention-guided masks allows the network to represent low-light and rain with different emphases before fusing them in a shared latent space (Lin et al., 2023).

The logic is operational rather than purely conceptual. Rain streaks are more visible in bright regions; low-light degradation dominates dark regions; the split feature streams are then concatenated and encoded so that subsequent layers can explicitly represent and differentiate the degradations. A plausible implication is that the latent representation becomes better aligned with the restoration task than a single undifferentiated degradation vector, because the model is encouraged to preserve degradation identity before feature fusion.

The paper’s description makes clear that DDR-Net is not a two-network ensemble in the usual sense. Its “dual” structure arises from region-conditioned decomposition of the same image, followed by shared representation learning. This differs from methods that use patches from arbitrary locations or assume that all local regions within one image should share a consistent degradation signature. The stated criticism of such assumptions is important for the low-light-rainy setting, where degradation features can vary greatly across locations (Lin et al., 2023).

4. Dual degradation loss and contrastive learning formulation

The learning signal for DDR-Net is a dual degradation loss formulated through contrastive learning. The paper explicitly rejects a conventional positive-pair definition based on patches from the same image, noting that in low-light-rainy images such patches may encode very different degradation states. Instead, the model uses multi-view learning anchored on augmentation invariance and clean-image separation (Lin et al., 2023).

The contrastive pair construction is:

• Positive pair: the feature from the original input $x$ and its augmented version $L^{2}$0, where augmentations include rotations and flips.
• Negative pair: the feature from the input $L^{2}$1 and the paired clean image $L^{2}$2.
• Shared encoder: all images are passed through the same encoder with parameter sharing.

The degradation loss is defined as

$L^{2}$3

where $L^{2}$4 is the P-Net encoder, $L^{2}$5 is the feature from the original image, $L^{2}$6 is the feature from the augmented image, and $L^{2}$7 is the feature from the paired clean image. The objective is to minimize the numerator and maximize the denominator, so that degradation features remain stable under simple augmentations but remain discriminative against the clean reference (Lin et al., 2023).

The full training objective combines degradation learning and restoration:

$L^{2}$8

with $L^{2}$9, $MAP$0, and $MAP$1 for the VGG perceptual loss. The restoration loss $MAP$2 is specified as $MAP$3 loss plus perceptual loss, and the paper also describes the total joint loss equivalently as $MAP$4 with carefully balanced weights (Lin et al., 2023).

5. Interface with the Restoration Network

DDR-Net’s output is not an end product; it is a guidance signal for the Restoration Network (R-Net). The paper states that the latent multi-channel degradation feature vector produced by P-Net is fed to R-Net as guidance, enabling the restoration network to focus on the actual, spatially dependent degradation factors (Lin et al., 2023).

R-Net is also described as U-Net-like. Its initial input is a concatenation of:

• the original low-light-rainy image, and
• a detail image, described as rain-reduced and derived via channel differences and the FFR-DG module.

The abstract further characterizes the restoration component through a Fourier Detail Guidance (FDG) module, which leverages near-rainless detailed images and focuses on texture details in both frequency and spatial domains. Mid-level skip connections and the latent codes from P-Net modulate R-Net at several decoder depths (Lin et al., 2023).

This division of responsibilities is structurally significant. DDR-Net isolates degradations into a latent code, while the restoration network uses both the raw observation and explicit degradation guidance. This suggests a restoration process conditioned on degradation identity rather than forced to infer all corruption factors implicitly from the image alone.

6. Training protocol, dataset, and empirical results

The training scheme uses both synthetic and real low-light rainy images. The reported setup includes random cropping to $MAP$5, Adam optimizer, batch size 24, and 200 epochs. Data augmentation by rotation and flipping is essential because it provides the positive pairs required by the degradation contrastive loss (Lin et al., 2023).

The work also contributes a Low-Light-Rainy (LLR) Dataset comprising:

• 8200 pairs synthetic images for training,
• 800 synthetic images for testing, and
• 430 real-world images for testing.

The synthetic images are generated by adjusting illumination of images from rain datasets so as to mimic realistic spatial distribution of rain and low-light, and real-world LLR images are collected and curated (Lin et al., 2023).

On the synthetic LLR benchmark, the paper reports the following retrained comparison:

Model	PSNR / SSIM	Params (M)
Zero-DCE*	21.75 / 0.7189	0.5
RCDNet*	25.04 / 0.8031	3.0
PyDiff*	29.06 / 0.8753	54.5
MIRNet*	29.31 / 0.8872	31.8
$MAP$6RIRNet (DDR-Net)	29.96 / 0.9007	11.9

The paper states that $MAP$7RIRNet achieves the highest PSNR and SSIM, with fewer parameters than several strong baselines, and also reports the fastest runtime versus most strong baselines. On real images, using the no-reference NIQE metric, DDR-Net records 13.881, compared with 14.658 for MIRNet* and 16.927 for the input image (Lin et al., 2023).

The ablation results attribute measurable gains to DDR-Net itself. Adding P-Net to R-Net yields an increase in PSNR of 0.58 dB over a single-branch degradation extractor, while adding the FFR-DG module further improves PSNR by 0.84 dB. The qualitative analysis in the paper reports that the model removes both rain and low-light artifacts while avoiding overexposure, underexposure, and rain residue that are common in cascade baselines (Lin et al., 2023).

7. Naming, scope, and potential ambiguity

Within this work, DDR-Net specifically means Dual Degradation Representation Network, and the paper also refers to it as P-Net. That designation is local to the low-light-rainy image restoration setting and should not be conflated with other arXiv uses of the same acronym (Lin et al., 2023).

The ambiguity is substantive rather than cosmetic. Separate papers use “DDR-Net” to denote a Dividing and Downsampling mixed Registration network for diffeomorphic image registration (Joshi et al., 2021), a Dynamic Depth Range Network for multi-view stereo (Yi et al., 2021), and a data-driven RetinaNet model for small object detection in aerial images (Tang et al., 3 Sep 2025). Those models solve unrelated problems and employ different architectural principles.

A common misconception, therefore, is to treat “DDR-Net” as a single canonical architecture. The literature represented here indicates instead that the acronym is overloaded across multiple research areas. In the context of joint deraining and low-light enhancement in the dark, the relevant meaning is the attention-guided, contrastively trained degradation representation module embedded in $MAP$8RIRNet (Lin et al., 2023).