AECR-Net: Compact Dehazing Deep Network

• AECR-Net is a compact, end-to-end autoencoder deep network for dehazing that integrates attention, contrastive regularization, adaptive mixup, and dynamic feature enhancement.
• It leverages a feature-attention backbone with U-Net skip connections and dynamic feature enhancement, achieving state-of-the-art PSNR and SSIM on both synthetic and real-world benchmarks.
• The model’s efficient design, with only 2.6–3M parameters, enables robust dehazing for adverse visibility conditions and potential applications in other image restoration tasks.

AECR-Net is a compact, end-to-end autoencoder-style deep network designed for single image dehazing and enhancement under challenging visibility conditions such as haze and smoke. The model integrates a feature-attention backbone, contrastive regularization, adaptive mixup, and a dynamic feature enhancement mechanism. AECR-Net has demonstrated state-of-the-art quantitative and qualitative results on both synthetic and real-world haze and smoke benchmarks, with practical utility further validated for gauge image interpretation in adverse visibility applications (Ramírez-Agudelo et al., 15 Jan 2026, Wu et al., 2021).

1. Architectural Components

At its core, AECR-Net employs an autoencoder architecture augmented with various feature fusion and regularization modules:

• Encoder Stem: Processes an RGB input $I_{in} \in \mathbb{R}^{3 \times H \times W}$ through a $7 \times 7$ convolution (64 channels), ReLU, and InstanceNorm, followed by downsampling (stride-2, 128 channels).
• Bottleneck: Comprises $N$ stacked AECR-Blocks, each integrating FFA-Net's Feature Attention (FA) block—composed of Channel-Attention (CA) and Pixel-Attention (PA) sub-modules—with local residuals: $X_{out} = X_{in} + FA(X_{in})$. Every $M$ blocks (typically $M = 2$), attributes a Dynamic Feature Enhancement (DFE) operation to adaptively re-scale channel statistics.
• Decoder: Mirrors the encoder with upsampling (using nearest-neighbor or transposed convolutions), ReLU, InstanceNorm, and final $3 \times 3$ convolution projecting to $3$ output channels with sigmoid output activation.
• Skip Connections: Employs U-Net-style skip connections to link encoder and decoder layers at corresponding spatial resolutions, enhancing feature fusion.
• Adaptive Mixup Module: During training only, generates synthetic feature-space interpolations (mix-of-pairs) between clean and hazy encodings to augment representation diversity for contrastive learning.

The model contains approximately $2.6$–$3$ million parameters, making it compact relative to other state-of-the-art dehazing networks (Wu et al., 2021).

2. Mathematical Formulation and Losses

AECR-Net’s objective combines image reconstruction fidelity with a pixel-level contrastive constraint:

• Reconstruction Loss: Per-pixel $\ell_1$ distance between dehazed output $\hat{Y}$ and ground truth $Y$:

$L_{rec}(\hat{Y}, Y) = \|Y - \hat{Y}\|_1$

• Contrastive Loss (InfoNCE): Given representations $z^+, z^-, \hat{z}$ extracted via a pre-trained feature extractor (e.g., VGG-19), AECR-Net enforces:

$$L_{ctr}(\hat{z}_i) = -\log \frac{\exp(\text{sim}(\hat{z}_i, z^+_i)/\tau)}{\sum_{j=1}^K \exp(\text{sim}(\hat{z}_i, \hat{z}_j)/\tau)}$$

where $\text{sim}(a,b)$ denotes cosine similarity and $\tau$ is a temperature hyperparameter.

• Total Objective: Weighted combination,

$L_{total} = L_{rec} + \lambda L_{ctr}$

with typical weighting $\lambda=1.0$ or, in some variants, $\beta=0.1$ (Wu et al., 2021).

Contrastive regularization constrains network outputs to lie closer to ground truth features while diverging from hazy inputs, narrowing the feasible restoration manifold (Wu et al., 2021).

3. Feature Engineering: Attention, Mixup, DFE

• Feature Attention (FA): Inherits both channel and pixel attention from FFA-Net to emphasize discriminative statistics at multiple representation levels.
• Adaptive Mixup: Parameterizes skip-level feature fusion via two learned interpolations:

$$f_{\uparrow 2} = \sigma(\theta_1)f_{\downarrow 1} + (1-\sigma(\theta_1))f_{\uparrow 1}$$

blending encoder and decoder activations to preserve spatial detail and yield sharper reconstructions.

• Dynamic Feature Enhancement (DFE): Utilizes two stacked modulated deformable convolutions:

$$y(p) = \sum_{k=1}^K w_k [x(p + p_k + \Delta p_k)] m_k$$

where $\Delta p_k$ are learned offsets, $m_k$ are modulation masks, and $w_k$ are kernel weights, adaptively expanding the receptive field.

Ablation experiments demonstrate each module’s contribution: DFE alone yields $\sim$+1.7$$dB <a href="https://www.emergentmind.com/topics/peak-signal-to-noise-ratio-psnr" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">PSNR</a>, adaptive mixup provides$$\sim$+0.2$ dB, and full contrastive regularizer delivers $\sim$+0.7$$dB over positives-only (<a href="/papers/2104.09367" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Wu et al., 2021</a>).</p> <h2 class='paper-heading' id='training-protocols-and-evaluation'>4. Training Protocols and Evaluation</h2> <p><strong>Datasets:</strong></p> <ul> <li>For gauge image enhancement, synthetic data is generated in Unreal Engine 5.1.1 with realistic global illumination, custom 3D gauge meshes, and both Exponential Height Fog and GPU-simulated smoke particles. Each scene yields images across 10 haze and 10 smoke density levels, with one clear reference.</li> <li>RESIDE benchmark is used for generic single-image dehazing, with SOTS (indoor) test set for evaluation, and Dense-Haze and NH-HAZE for real-world validation.</li> </ul> <p><strong>Hyperparameters:</strong></p> <ul> <li>Optimizer: Adam$$(\beta_1=0.9, \beta_2=0.999, \epsilon=10^{-8})$</li> <li>Initial LR:$10^{-4}$or$2 \times 10^{-4}$$, stepped or cosine annealing</li> <li>Batch size:$$4$(gauge) or$16$(RESIDE)</li> <li>Epochs:$100$</li> <li>Loss weights:$\lambda=1.0$,$\beta=0.1$</li> <li>Feature extraction for contrastive loss layers: VGG-19, layers indexed {1,3,5,9,13} with progressively increasing weights</li> </ul> <p>No data augmentations are applied beyond density-level variation.</p> <h2 class='paper-heading' id='quantitative-performance'>5. Quantitative Performance</h2><div class='overflow-x-auto max-w-full my-4'><table class='table border-collapse w-full' style='table-layout: fixed'><thead><tr> <th>Dataset</th> <th>Method</th> <th>PSNR (dB)</th> <th>SSIM</th> <th>Parameters</th> </tr> </thead><tbody><tr> <td>Haze (Gauge)</td> <td>AECR-Net</td> <td>∼44</td> <td>0.98</td> <td>3M</td> </tr> <tr> <td></td> <td>FFA-Net</td> <td>∼30</td> <td>0.96</td> <td>-</td> </tr> <tr> <td></td> <td>BCCR</td> <td>∼12</td> <td>0.65</td> <td>-</td> </tr> <tr> <td>Smoke (Gauge)</td> <td>AECR-Net</td> <td>∼37</td> <td>0.96</td> <td>3M</td> </tr> <tr> <td></td> <td>FFA-Net</td> <td>∼26</td> <td>0.94</td> <td>-</td> </tr> <tr> <td></td> <td>BCCR</td> <td>∼9</td> <td>0.55</td> <td>-</td> </tr> <tr> <td>RESIDE/SOTS</td> <td>AECR-Net</td> <td>37.17</td> <td>0.990</td> <td>2.6M</td> </tr> <tr> <td></td> <td>FFA-Net</td> <td>36.39</td> <td>0.989</td> <td>4.68M</td> </tr> <tr> <td></td> <td>(Others)</td> <td>∼30</td> <td>∼0.97</td> <td>&gt;3M</td> </tr> </tbody></table></div> <p>On both synthetic and real-world datasets, AECR-Net matches or outperforms prior methods in PSNR and SSIM, despite its low model size. In custom smoke and haze gauge datasets, AECR-Net improves PSNR by roughly $+13$$dB over FFA-Net (<a href="/papers/2601.10537" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Ramírez-Agudelo et al., 15 Jan 2026</a>, <a href="/papers/2104.09367" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Wu et al., 2021</a>).</p> <h2 class='paper-heading' id='component-level-analysis-and-practical-impact'>6. Component-Level Analysis and Practical Impact</h2> <ul> <li><strong>Contrastive Regularization</strong>: Adding$$L_{ctr}$increases average PSNR by about$+2$dB (gauge enhancement) and$+0.7$$dB (RESIDE SOTS).</li> <li><strong>Dynamic Feature Enhancement</strong>: Disabling this module leads to$$\sim$0.01 drop in SSIM and $\sim$1.7 dB PSNR loss.

Adaptive Mixup: Reduces overfitting to mid-level densities, improving generalization to unseen visibility conditions.

FA-Block Depth: Increasing the number of FA blocks to $N=8$$improves performance, with diminishing returns beyond$$N>8$.

Parameter Efficiency: AECR-Net’s compactness (2.6–3M parameters) provides substantial computational and memory advantages relative to previous deep dehazing models (Wu et al., 2021).

In infrastructure and emergency response settings, AECR-Net’s enhanced outputs enable more robust post-processing pipelines for automatic and autonomous gauge interpretation, critical in haze- and smoke-obscured environments (Ramírez-Agudelo et al., 15 Jan 2026).

7. Availability and Extensions

The official implementation of AECR-Net is available at https://github.com/GlassyWu/AECR-Net. Contrastive regularization has shown universality: adding CR to alternative single-image dehazing architectures yields consistent PSNR and SSIM gains without increased inference costs, supporting AECR-Net’s utility as a general model design motif.

A plausible implication is that AECR-Net’s architecture and learning protocol may generalize to deblurring, low-light enhancement, and other single-image restoration tasks where information degradation results from complex, content-dependent imaging perturbations.

---

Key References:

• "Enhancing the quality of gauge images captured in smoke and haze scenes through deep learning" (Ramírez-Agudelo et al., 15 Jan 2026)
• "Contrastive Learning for Compact Single Image Dehazing" (Wu et al., 2021)