Source-linked AI summary
Cycle-Dehaze: Enhanced CycleGAN for Single Image Dehazing
Deniz Engin, Anıl Genç, Hazım Kemal Ekenel
TL;DR
Single-image dehazing must restore visibility despite limited training information and atmospheric-model dependencies. Cycle-Dehaze extends CycleGAN with perceptual and cycle-consistency losses and Laplacian-pyramid upscaling, producing improved results over CycleGAN while cross-dataset color recovery remains weaker.
Problem
Single-image dehazing seeks visibility enhancement without paired hazy-ground-truth images or atmospheric scattering-model parameters.
Method
Cycle-Dehaze enhances CycleGAN with cyclic perceptual-consistency loss and uses Laplacian-pyramid upscaling after low-resolution dehazing.
Results
Cycle-Dehaze achieves higher PSNR and SSIM values than CycleGAN on NYU-Depth and reports quantitative and qualitative improvements overall.
Takeaways & Limitations
The approach supports unpaired, atmospheric-model-independent dehazing across NYU-Depth, I-HAZE, and O-HAZE, including cross-dataset evaluation.
Takeaways & Limitations
In cross-dataset settings, color recovery is worse and PSNR is lower because thick haze can prevent estimating the actual ground-truth color.
Abstract
from arXiv · showhide
In this paper, we present an end-to-end network, called Cycle-Dehaze, for single image dehazing problem, which does not require pairs of hazy and corresponding ground truth images for training. That is, we train the network by feeding clean and hazy images in an unpaired manner. Moreover, the proposed approach does not rely on estimation of the atmospheric scattering model parameters. Our method enhances CycleGAN formulation by combining cycle-consistency and perceptual losses in order to improve the quality of textural information recovery and generate visually better haze-free images. Typically, deep learning models for dehazing take low resolution images as input and produce low resolution outputs. However, in the NTIRE 2018 challenge on single image dehazing, high resolution images were provided. Therefore, we apply bicubic downscaling. After obtaining low-resolution outputs from the network, we utilize the Laplacian pyramid to upscale the output images to the original resolution. We conduct experiments on NYU-Depth, I-HAZE, and O-HAZE datasets. Extensive experiments demonstrate that the proposed approach improves CycleGAN method both quantitatively and qualitatively.
1. Introduction
Single-image dehazing improves visibility for computer-vision applications, but existing approaches may require paired data or atmospheric-model parameters. Cycle-Dehaze enhances CycleGAN with perceptual consistency and Laplacian upscaling for unpaired, high-resolution dehazing.
- Haze reduces scene visibility and can hinder object detection, tracking, and segmentation, motivating dehazing as preprocessing.
- Experiments on NYU-Depth, I-HAZE, and O-HAZE include cross-dataset scenarios demonstrating the model's generalizability.
- The method enhances CycleGAN by adding cyclic perceptual-consistency loss alongside cycle-consistency loss to recover textural information.
- Cycle-Dehaze trains on hazy and clean images without paired correspondences or atmospheric scattering-model parameters.
- Laplacian-pyramid upscaling reconstructs high-resolution dehazed images after bicubic downscaling and low-resolution network processing.
2. Related Work
Related single-image dehazing work includes physical-model priors and learning-based methods. These approaches differ in whether they estimate atmospheric parameters, use additional scene information, or generate clean images directly.
- Single-image dehazing commonly estimates atmospheric scattering-model parameters, including scene radiance, atmospheric light, and transmission.
- Multiple-image and polarization-based methods use changing atmospheric conditions or filters, but their applicability is constrained by acquisition requirements and scene assumptions.
- Prior-based methods use image statistics such as dark-channel, color-attenuation, and haze-line priors to estimate haze-related quantities.
- Learning-based methods use CNNs or GANs to estimate transmission and atmospheric light, recover clean images, or produce haze-free images directly.
3. Proposed Method
Cycle-Dehaze enhances CycleGAN for single-image dehazing by combining cycle-consistency with perceptual consistency and using Laplacian upscaling for high-resolution outputs. Its architecture uses two generators and two discriminators, with perceptual features preserving image structure and texture.
- Cyclic perceptual-consistency loss: Cycle-Dehaze extends CycleGAN with cyclic perceptual-consistency loss to improve texture preservation and visual quality during single-image dehazing.The loss compares original and reconstructed cyclic images in feature space as well as pixel space.
- Network architecture: The network contains two generators, G and F, and two discriminators, Dx and Dy, for hazy-to-clean and clean-to-hazy translation.
- Full objective: The full objective adds a weighted cyclic perceptual loss to the original CycleGAN objective, with γ controlling its effect.The added loss is optimized alongside the CycleGAN generators, discriminators, and cycle-consistency terms.
- Laplacian upscaling: Laplacian upscaling replaces the pyramid’s top layer with the 256 × 256 dehazed output to reconstruct a high-resolution image while preserving edges.The network processes 256 × 256 images because of GPU limitations; upscaling is optional for high-resolution images.
- Implementation details: Training uses TensorFlow and Adam with learning rate 1e−4, approximately 40 epochs per dataset, and testing takes about 8 seconds per image on an Intel Core i7-5820K CPU.
4. Experiments and Results
Experiments on NYU-Depth, I-HAZE, O-HAZE, natural images, and cross-dataset settings show that Cycle-Dehaze improves CycleGAN quantitatively and qualitatively while addressing high-resolution inputs.
- Experimental setup: Experiments covered NYU-Depth, I-HAZE, O-HAZE, natural hazy images, and cross-dataset testing between NYU-Depth and I-HAZE.NYU-Depth contains 1,449 clean–hazy pairs; I-HAZE and O-HAZE contain 25 indoor and 35 outdoor hazy images with ground truth, respectively.
- NYU-Depth results: Cycle-Dehaze achieved the second-best PSNR on NYU-Depth and exceeded CycleGAN in both PSNR and SSIM.The authors report nearly similar PSNR to methods using atmospheric-scattering-model parameters.
- I-HAZE and O-HAZE results: Cycle-Dehaze produced higher PSNR and SSIM than CycleGAN on every I-HAZE and O-HAZE challenge track.Outdoor scenes had lower SSIM but higher PSNR than indoor scenes, which the authors relate to scene distance and fog distribution.
- I-HAZE and O-HAZE results: Qualitatively, Cycle-Dehaze generated less noise and sharper edges than CycleGAN, although repeated outdoor textures reduced edge sharpness.The authors attribute noise reduction to cyclic perceptual-consistency loss and sharper edges to the Laplacian pyramid.
- Cross-dataset and natural-image results: Cross-dataset testing yielded high PSNR and SSIM, with results as good as CycleGAN’s regular single-dataset testing and evidence of learning dehazing beyond dataset-specific overfitting.On natural images, the method preserved neighboring color tones, shadows, and depth more perceptibly.
5. Conclusion
Cycle-Dehaze directly produces haze-free images without atmospheric-model estimation or paired training data. It combines perceptual and cycle-consistency losses, uses Laplacian-pyramid upscaling, and reports better image quality and generalizability than CycleGAN.
- Cycle-Dehaze directly generates haze-free images without estimating atmospheric scattering parameters or requiring paired hazy and ground-truth images.
- Combining perceptual loss with CycleGAN’s cycle-consistency loss is intended to retain high visual quality and improve textural information recovery.
- Laplacian-pyramid upscaling reduces distortion when restoring low-resolution network outputs to the original image resolution.
- The experiments report visually better images and higher PSNR and SSIM values than CycleGAN.
- Additional cross-dataset experiments demonstrate generalizability across different domains.