[Paper Review] An All-in-One Network for Dehazing and Beyond

Boyi Li, Xiulian Peng+3 more|arXiv (Cornell University)|Jul 20, 2017

Image Enhancement Techniques36 references136 citations

TL;DR

AOD-Net directly regresses haze-free images from hazy inputs with an end-to-end CNN grounded in a reformulated atmospheric model, and it can be jointly trained with high-level tasks (e.g., Faster R-CNN) to boost object detection on hazy images.

ABSTRACT

This paper proposes an image dehazing model built with a convolutional neural network (CNN), called All-in-One Dehazing Network (AOD-Net). It is designed based on a re-formulated atmospheric scattering model. Instead of estimating the transmission matrix and the atmospheric light separately as most previous models did, AOD-Net directly generates the clean image through a light-weight CNN. Such a novel end-to-end design makes it easy to embed AOD-Net into other deep models, e.g., Faster R-CNN, for improving high-level task performance on hazy images. Experimental results on both synthesized and natural hazy image datasets demonstrate our superior performance than the state-of-the-art in terms of PSNR, SSIM and the subjective visual quality. Furthermore, when concatenating AOD-Net with Faster R-CNN and training the joint pipeline from end to end, we witness a large improvement of the object detection performance on hazy images.

Motivation & Objective

• Motivate robust single-image dehazing without explicit transmission and atmospheric light estimation.
• Propose an end-to-end CNN that outputs haze-free images via a reformulated atmospheric model.
• Design a lightweight, input-adaptive network that jointly estimates haze parameters and recovers the clean image.
• Demonstrate superior restoration quality (PSNR/SSIM) and improved high-level task performance on hazy images.

Proposed method

• Reformulate the atmospheric scattering model to combine t(x) and A into a single input-dependent parameter K(x).
• Develop a two-module network: a K-estimation module with multi-scale convolutions and a clean image generation module using K(x) to Produce J(x).
• Use a lightweight architecture with five convolutional layers in the K-estimation module and minimal filters (three per layer).
• Train end-to-end with Mean Squared Error loss to optimize direct reconstruction of J(x) from I(x).
• Show that joint estimation of t(x) and A within K(x) yields better illumination and texture fidelity than a naive two-step approach.
• Demonstrate the ability to integrate AOD-Net into higher-level models (e.g., Faster R-CNN) and train jointly (JAOD-Faster R-CNN) for improved detection on hazy images.

Experimental results

Research questions

• RQ1 Does an end-to-end dehazing model that directly regresses the clean image outperform two-step transmission-based methods?
• RQ2 Can a reformulated atmospheric model with a single adaptive parameter improve restoration quality (PSNR/SSIM) and visual fidelity?
• RQ3 Does joint training of dehazing with a high-level task (e.g., object detection) improve performance on hazy images?
• RQ4 Is a lightweight, input-adaptive network sufficient for robust dehazing across synthetic and natural hazy scenes?

Key findings

• AOD-Net achieves higher SSIM and competitive PSNR compared with state-of-the-art methods on synthetic hazy datasets (TestSet A and TestSet B).
• AOD-Net provides more faithful global illumination and texture details, reflected in higher SSIM and lower mean-image MSE components.
• The joint JAOD-Faster R-CNN pipeline yields substantial gains in object detection mAP under haze compared to naive or non-joint approaches.
• AOD-Net runs much faster (0.65–0.65 seconds per image on CPU/GPU benchmarks) than several baselines, enabling efficient deployment.
• The end-to-end design enables restoration to be tuned for downstream tasks, outperforming two-stage pipelines in dense haze scenarios.
• Dehazing can improve performance on natural hazy images and remains robust when applied to haze-free inputs.