Source-linked AI summary
Contrastive Learning for Compact Single Image Dehazing
Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, Lizhuang Ma
TL;DR
Single image dehazing suffers from information degeneration, while existing methods underuse hazy negative information and often incur high computation and memory costs. The paper proposes AECR-Net, combining contrastive regularization with a compact autoencoder-like network, and reports superior performance over state-of-the-art methods on synthetic and real-world datasets.
Problem
Existing dehazing methods mainly use clear images as positive samples, leave negative information unexploited, and often increase depth and width at substantial computation and memory cost.
Method
AECR-Net combines contrastive regularization, which pulls restored images toward clear images and pushes them away from hazy images, with a compact autoencoder-like network using adaptive mixup and dynamic feature enhancement.
Results
AECR-Net demonstrates superior performance gains over state-of-the-art methods on synthetic and real-world datasets.
Takeaways & Limitations
Contrastive regularization exploits both positive and negative image information, while the compact architecture targets the trade-off between dehazing performance and parameter cost.
Takeaways & Limitations
Transmission-map and atmospheric-light estimation methods face cumulative reconstruction errors, and obtaining real-world ground truth for these quantities is difficult or expensive.
Abstract
from arXiv · showhide
Single image dehazing is a challenging ill-posed problem due to the severe information degeneration. However, existing deep learning based dehazing methods only adopt clear images as positive samples to guide the training of dehazing network while negative information is unexploited. Moreover, most of them focus on strengthening the dehazing network with an increase of depth and width, leading to a significant requirement of computation and memory. In this paper, we propose a novel contrastive regularization (CR) built upon contrastive learning to exploit both the information of hazy images and clear images as negative and positive samples, respectively. CR ensures that the restored image is pulled to closer to the clear image and pushed to far away from the hazy image in the representation space. Furthermore, considering trade-off between performance and memory storage, we develop a compact dehazing network based on autoencoder-like (AE) framework. It involves an adaptive mixup operation and a dynamic feature enhancement module, which can benefit from preserving information flow adaptively and expanding the receptive field to improve the network's transformation capability, respectively. We term our dehazing network with autoencoder and contrastive regularization as AECR-Net. The extensive experiments on synthetic and real-world datasets demonstrate that our AECR-Net surpass the state-of-the-art approaches. The code is released in https://github.com/GlassyWu/AECR-Net.
1. Introduction
The paper identifies limitations in positive-only dehazing supervision and proposes contrastive regularization alongside a compact autoencoder-like network. AECR-Net uses positive and negative image information while targeting a favorable performance–parameter trade-off.
- Haze degrades object appearance and contrast and can impair high-level tasks such as object detection and scene understanding.
- Existing dehazing methods rely mainly on clear images as positive samples, while reconstruction losses can leave image details insufficiently handled and produce color distortion.
- Positive-image regularization provides an upper bound, but unexploited negative-image information can still allow artifacts or unsatisfactory results.
- Contrastive regularization pulls restored images toward clear images and pushes them away from hazy images in representation space.
- AECR-Net combines contrastive regularization with a compact autoencoder-like network designed for a favorable performance–parameter trade-off.
- Adaptive mixup preserves information flow adaptively, while dynamic feature enhancement improves the network’s transformation capability.
2. Related Work
Related work includes prior-based and learning-based dehazing methods, with limitations involving scene-specific assumptions, estimation errors, and expensive supervision. The paper addresses these issues with contrastive regularization and a compact autoencoder-like architecture.
- 2. Related Work: Single image dehazing methods are categorized into prior-based approaches and learning-based approaches.
- 2.1. Single Image Haze Removal: Prior-based methods use physical scattering models and handcrafted priors, but their scene-dependent assumptions reduce robustness in complex practical scenes.
- 2.1. Single Image Haze Removal: Transmission-map and atmospheric-light estimation can accumulate errors, while collecting corresponding real-world ground truth is difficult or expensive.
- 2.1. Single Image Haze Removal: End-to-end methods often use clear images as positive supervision, yet artifacts can remain because negative-image information is not exploited.
- 2.1. Single Image Haze Removal: The paper combines contrastive regularization using positive and negative images with a compact autoencoder-like network that reduces layers and spatial size.
- 2.2. Contrastive Learning: Contrastive learning pulls anchors toward positive points and pushes them away from negative points in representation space.
3. Our Method
The method combines a compact autoencoder-like dehazing network with adaptive feature fusion, dynamic feature enhancement, and contrastive regularization. These components reduce computation while preserving information, expanding feature context, and constraining restored images toward clear images and away from hazy inputs.
- 3.2. Autoencoder-like Dehazing Network: The AE-like network performs dense convolution in low-resolution space and reduces the number of layers to trade performance against parameters.It uses downsampling before dense feature processing and upsampling to reconstruct the output.
- 3.2.1 Adaptive Mixup for Feature Preserving: Adaptive mixup learns to fuse shallow downsampling features with high-level upsampling features, preserving information flow through the network.The fusion uses learnable factors determined by sigmoid-transformed parameters rather than fixed coefficients.
- 3.2.2 Dynamic Feature Enhancement: Dynamic feature enhancement uses a flexible receptive field to capture structured information that fixed grid kernels may miss.The design is motivated by the limited receptive field of fixed kernels and the potential gridding artifacts of dilated convolutions.
- 3.3. Contrastive Regularization: Contrastive regularization pairs restored and clear images positively while pairing restored and hazy images negatively in a shared feature space.Features are extracted from multiple layers of a fixed pretrained model, and the contrastive term complements reconstruction loss.
- 3.3. Contrastive Regularization: The overall dehazing objective combines L1 reconstruction with contrastive regularization, balanced by a hyperparameter and optimized end to end.The reconstruction term aligns the restored image with its ground truth, while contrastive regularization pulls it toward clear images and pushes it from hazy images.
4. Experiments
Experiments evaluate AECR-Net on synthetic and real-world dehazing datasets using PSNR and SSIM, comparing its performance, compactness, ablations, and contrastive-regularization variants against prior methods.
- Experiment Setup: Experiments use RESIDE SOTS for synthetic evaluation and Dense-Haze and NH-HAZE for real-world evaluation, with PSNR and SSIM as metrics.The comparisons include prior-based, physical-model-based, and hazy-to-clear translation methods.
- Comparison with State-of-the-art Methods: 37.17dB PSNR and 0.9901 SSIM make AECR-Net the best-performing method on the RESIDE SOTS synthetic dataset.It gains 0.78dB PSNR and 0.0015 SSIM over FFA-Net while using 2M fewer parameters.
- Comparison with State-of-the-art Methods: Visual comparisons show AECR-Net producing natural restored images, while competing methods exhibit color distortion, texture loss, residual haze, or artifacts.These comparisons cover SOTS, Dense-Haze, and NH-HAZE visual results.
- Comparison with State-of-the-art Methods: 19.88dB PSNR and 0.7173 SSIM are achieved on NH-HAZE, while AECR-Net records the highest 15.80dB PSNR on Dense-Haze.On NH-HAZE, AECR-Net outperforms the compared methods; on Dense-Haze, MSBDN has about 0.02 higher SSIM but uses 12× more parameters.
- Ablation Study: Adaptive mixup improves PSNR by 0.19dB and 0.7dB in two base-network comparisons and by 0.61dB over skip connection.The operation adds negligible parameters and adaptively fuses different features.
- Ablation Study: DFE improves PSNR by 1.65dB from base to base+DFE and by 2.16dB from base+mixup to base+DFE+mixup.The deeper-layer placement performs better than shallow-layer placement.
- Universal Contrastive Regularization: CR with both negative and positive samples substantially improves the AECR-Net, while CR* using only positive samples yields only 0.26dB PSNR and 0.002 SSIM gains.Applying CR to several state-of-the-art methods also improves their performance without increasing inference parameters.
5. Conclusion
AECR-Net combines contrastive regularization with a compact autoencoder-like dehazing network, achieving superior performance gains over state-of-the-art methods on synthetic and real-world datasets.
- AECR-Net combines contrastive regularization with an autoencoder-like dehazing network.The network uses contrastive regularization alongside a compact AE-like architecture.
- Contrastive regularization pulls restored images toward clear images and pushes them away from hazy images in representation space.This uses clear and hazy images as positive and negative references, respectively.
- The compact AE-like network adaptively preserves information flow and expands the receptive field through adaptive mixup and dynamic feature enhancement.These components improve the network’s transformation capability.
- AECR-Net demonstrates superior performance gains over state-of-the-art methods on synthetic and real-world datasets.The paper evaluates the method across both synthetic and real-world dehazing data.