💻 computer science

Dual-Path Learning based on Frequency Structural Decoupling and Regional-Aware Fusion for Low-Light Image Super-Resolution

This paper proposes the Decoupling then Perceive (DTP) framework, a novel frequency-aware approach for Low-Light Image Super-Resolution that employs Frequency-aware Structural Decoupling, Semantics-specific Dual-path Representation, and Cross-frequency Semantic Recomposition to independently model luminance and texture, thereby significantly outperforming state-of-the-art methods in PSNR, SSIM, and LPIPS metrics.

Original authors: Ji-Xuan He, Jia-Cheng Zhao, Feng-Qi Cui, Jinyang Huang, Yang Liu, Sirui Zhao, Meng Li, Zhi Liu

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ji-Xuan He, Jia-Cheng Zhao, Feng-Qi Cui, Jinyang Huang, Yang Liu, Sirui Zhao, Meng Li, Zhi Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to restore an old, faded, and blurry photograph taken in a pitch-black room. You want to make it bright enough to see clearly (Low-Light Enhancement) and sharp enough to see the tiny details like the texture of a shirt or the text on a sign (Super-Resolution).

Doing both at once is incredibly hard. Most existing methods try to fix the brightness and the sharpness simultaneously, like trying to paint a picture while simultaneously fixing the canvas. The result? The brightness adjustments often smear the details, and the attempts to sharpen the image make the shadows look weird and noisy. It's a messy compromise.

This paper introduces a new method called DTP (Decoupling then Perceive). Think of it as a smart, two-step kitchen process for restoring a ruined meal, rather than trying to fix everything in one chaotic pot.

Here is how it works, broken down into simple analogies:

1. The Problem: The "Smoothie" Mistake

Imagine you have a smoothie made of ice cream (brightness) and crunchy nuts (texture). If you try to fix the flavor by blending them all together, you can't taste the nuts anymore, and the ice cream gets weirdly icy.

Old Methods: They treat the image like that smoothie. They try to brighten and sharpen everything at once, causing the "nuts" (details) to get mushy and the "ice cream" (light) to spill over into places it shouldn't.

2. The Solution: The "Decoupling" (Separating the Ingredients)

The first step of DTP is Frequency-Aware Structural Decoupling (FSD).

The Analogy: Instead of blending, this step acts like a magic sieve. It separates the image into two distinct piles:
- Pile A (Low-Frequency): This is the "Luminance" or the overall brightness and shadows. It's the smooth, soft part of the image.
- Pile B (High-Frequency): This is the "Texture." It's the sharp edges, the noise, the fine lines, and the details.
Why it helps: By separating them, the computer can stop the brightness from smearing the details. It's like putting the ice cream in one bowl and the nuts in another so you can fix each one perfectly without ruining the other.

3. The Specialization: The "Dual-Path" Chefs

Once the ingredients are separated, DTP sends them to two different "chefs" (branches) who specialize in exactly what they need to do. This is the Semantics-specific Dual-path Representation (SDR).

Chef A (The Luminance Chef): This chef only looks at the brightness pile. They use a "Bio-inspired" recipe (mimicking how human eyes adjust to light) to gently brighten the dark areas without blowing out the highlights. They make sure the light looks natural, like a sunrise, not a flashbulb.
Chef B (The Texture Chef): This chef looks at the texture pile. Their job is to clean up the noise (like dust on a lens) and sharpen the edges. They use a "Residual" technique, which is like a sculptor chipping away the bad stone to reveal the perfect shape underneath, ensuring the details are crisp but not jagged.

4. The Reunion: The "Smart Fusion"

Now that the brightness is perfect and the texture is sharp, they need to be put back together.

The Old Way: Just dumping the two bowls back together (naive mixing).
The DTP Way: This uses a Cross-frequency Semantic Recomposition (CSR) module. Think of this as a Master Sommelier.
- The Sommelier doesn't just mix the bowls; they taste the result and decide exactly how much of the brightness and how much of the texture should go into the final glass.
- They use "Attention" (like a spotlight) to focus on specific areas. If a part of the image needs more light, they add more brightness. If a part needs more detail, they add more texture.
- This ensures the final image looks natural, with the light and the details perfectly aligned.

The Result

When the authors tested this method, it was like comparing a blurry, washed-out photo to a high-definition, crystal-clear masterpiece.

Better Numbers: It improved the clarity (PSNR) and the "feel" of the image (LPIPS) significantly more than any other current method.
Extreme Conditions: Even in lighting so dark it's almost invisible (like -4.5 EV, which is incredibly dark), this method could still see the edges of a bicycle wheel or the text on a bottle, while other methods just gave up or created weird artifacts.

Summary

In short, DTP stops trying to fix a broken image with a sledgehammer. Instead, it:

Separates the light from the details.
Fixes the light and the details separately with specialized tools.
Recombines them intelligently so they fit together perfectly.

It's the difference between trying to fix a car engine while driving it versus taking the car apart, fixing each part on a workbench, and then putting it back together to run like new.

1. Problem Statement

Low-Light Image Super-Resolution (LLISR) is a critical task in computer vision aimed at restoring fine visual details and perceptual quality from images captured under insufficient illumination. However, existing methods face three fundamental challenges:

Semantic Entanglement: Traditional approaches often model luminance (brightness) and texture (details) jointly in a unified spatial feature space. This lack of explicit disentanglement causes the model to confuse luminance artifacts with true textures, leading to structural distortion and perceptual inconsistency.
Lack of Subspace Sensitivity: Most methods use flat or undifferentiated representations (e.g., shared channels) that fail to adaptively respond to the distinct frequency characteristics of luminance (low-frequency) and texture (high-frequency). This results in feature interference and insufficient specialization.
Static Fusion Limitations: Mainstream fusion strategies often rely on fixed weights or naive concatenation, ignoring regional variations in luminance and texture distribution. This rigidity leads to semantic misalignment and local structural degradation.

2. Methodology: Decoupling then Perceive (DTP)

The authors propose DTP, a novel frequency-aware framework that explicitly separates, models, and fuses heterogeneous visual representations. The workflow consists of three main stages:

A. Frequency-Aware Structural Decoupling (FSD)

Goal: To mitigate early-stage semantic entanglement by separating the input into distinct frequency subspaces.
Mechanism:
- Applies a learnable wavelet transform to decompose the low-light input into four subbands: Low-Low (LL, global luminance), and High-frequency bands (LH, HL, HH, textures).
- Uses a learnable vector to reweigh the stacked subbands and applies a KL divergence loss to ensure the luminance component remains compact.
- Explicitly separates the features into two independent branches: $I_{llr}$ (structure-preserving low-frequency luminance) and $T_{llr}$ (fine-grained high-frequency textures).

B. Semantics-Specific Dual-Path Representation (SDR)

Goal: To perform targeted enhancement on the decoupled components without cross-domain interference.
Luminance Branch (Bio-inspired Activation):
- Mimics human photoreceptor responses using a non-linear transformation based on the Naka–Rushton equation.
- Enhances brightness in dark regions while preventing overexposure and maintaining structural contrast.
- Formula: $I_{hlr} = \frac{(I_{llr})^\gamma}{(I_{llr})^\gamma + \sigma^\gamma} + \beta \cdot \frac{\mu_{in}}{\mu_{out} + \epsilon}$ .
Texture Branch (Hierarchical Denoising):
- Addresses the noise vulnerability of high-frequency textures in extreme low light.
- Employs a residual learning framework where noise is iteratively removed while preserving fine details: $T_i = T_{i-1} + R_i(T_{i-1})$ .

C. Cross-frequency Semantic Recomposition (CSR)

Goal: To adaptively fuse the enhanced luminance and restored texture into a unified high-fidelity output.
Mechanism:
- Instead of naive concatenation, it uses dual-branch attention (Spatial and Channel attention) to assess semantic contributions.
- A learnable gating function ( $G_{att}$ ) balances the fusion based on input dynamics, creating a soft selection mask.
- Includes a residual refinement path to reinforce high-frequency structure retention before the final upsampling and reconstruction.

3. Key Contributions

First Frequency-Aware Disentanglement in LLISR: The paper introduces the principle of explicitly decoupling luminance and texture in the frequency domain, aligning modeling pathways with their specific semantic roles to mitigate early-stage entanglement.
Novel FSD Module: Proposes an adaptive transformation scheme to separate inputs into distinct frequency subspaces, enabling compact representations for targeted modeling.
Dual-Path Modeling & Adaptive Fusion: Introduces the SDR (using bio-inspired adjustment and residual refinement) and CSR (using spatial-channel attention) modules. This design improves regional generalization and structural fidelity.
State-of-the-Art Performance: Demonstrates consistent superiority over existing methods across diverse low-light scenarios and scaling factors.

4. Experimental Results

The method was evaluated on the RELLISUR dataset (a benchmark for low-light super-resolution) at $\times2$ and $\times4$ scales.

Quantitative Performance:
- Compared to the best existing SOTA algorithm, DTP achieved:
  - +1.6% improvement in PSNR.
  - +9.6% improvement in SSIM.
  - -48% reduction in LPIPS (indicating significantly better perceptual quality).
- On the $\times2$ scale, DTP achieved 23.15 PSNR and 0.816 SSIM, outperforming the runner-up (BrZoNet) by a significant margin.
Qualitative Analysis:
- Visual comparisons show DTP produces sharper structural boundaries and clearer text/labels compared to methods that suffer from over-smoothing or noisy over-enhancement.
- Robustness: Under extreme low-light conditions (down to -4.5 EV), DTP maintained geometric layout and spoke sharpness, whereas baseline methods exhibited structure collapse or oversaturation.
Ablation Studies: Confirmed that each module (FSD, SDR, CSR) contributes positively, with the full framework yielding the best results. The combination of FSD and SDR was particularly effective in resolving semantic confusion.

5. Significance

This paper addresses a critical bottleneck in low-light image restoration: the trade-off between brightness enhancement and detail preservation. By shifting from spatial-domain joint modeling to frequency-domain structural decoupling, DTP offers a more robust and interpretable solution.

Theoretical Impact: It establishes a new paradigm for handling dual-degradation tasks (low-light + low-resolution) by treating luminance and texture as semantically independent components.
Practical Application: The method is highly relevant for applications requiring high fidelity in dark environments, such as night surveillance, autonomous driving, and low-light photography, where preserving both global illumination and fine textures is essential.
Efficiency: Unlike other complex fusion strategies, DTP achieves superior performance through a streamlined pipeline of decoupling, specialization, and recomposition, avoiding unnecessary computational overhead while solving underlying semantic interference.