PaperlayerView exact revision on arXiv

Source-linked AI summary

An All-in-One Network for Dehazing and Beyond

Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, Dan Feng

arXiv:1707.06543v1cs.CVcs.AI

TL;DR

Dehazing methods traditionally estimate transmission and atmospheric light separately, while haze also harms downstream vision tasks. The paper introduces AOD-Net, an end-to-end CNN that directly reconstructs clean images and can be jointly optimized with Faster R-CNN. Across synthetic and natural haze evaluations, it reports stronger restoration quality and improved hazy-image object detection, including a heavy-haze mAP increase from 0.5794 to 0.6819.

  • Problem

    Existing dehazing methods separately estimate physical parameters, and the relationship between dehazing quality and high-level vision performance had not been quantitatively studied.

  • Method

    AOD-Net reformulates the atmospheric scattering model into a lightweight end-to-end CNN that directly reconstructs clean images and can be jointly optimized with Faster R-CNN.

  • Results

    AOD-Net outperforms several state-of-the-art methods in PSNR, SSIM, and visual quality, while joint tuning improves hazy-image object detection.

  • Takeaways & Limitations

    The all-in-one design supports unified restoration and recognition pipelines, with JAOD-Faster R-CNN outperforming retrained Faster R-CNN across haze conditions.

Abstract

from arXiv · show

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.

I. INTRODUCTION

Haze degrades visibility and complicates image restoration, while existing dehazing methods separately estimate physical parameters and face limited evaluation evidence, especially for natural images.

  • Haze reduces outdoor-image contrast and weakens surface colors, while also jeopardizing object detection and recognition.
  • Prior work used physical priors such as transmission estimation and atmospheric-light recovery to remove haze.
  • End-to-end dehazing had not directly regressed clean images from hazy inputs, partly because haze introduces depth-dependent, signal-dependent noise.
  • Existing dehazing methods estimate transmission and atmospheric light separately before reconstructing the clean image.This procedure can accumulate or amplify errors and produce sub-optimal restoration quality.

2) Missing Link with High-Level Vision Tasks:

The paper identifies a missing connection between dehazing and high-level vision tasks: restoration is usually treated as preprocessing, but its effect on detection had not been explored quantitatively.

  • 2) Missing Link with High-Level Vision Tasks:: Image restoration is commonly used as preprocessing for mid-level and high-level vision tasks.
  • 2) Missing Link with High-Level Vision Tasks:: Degradations can deteriorate object detection and recognition, yet prior work had not connected dehazing results with high-level task performance.
  • C. Main Contributions: AOD-Net directly produces clean images from hazy inputs through an end-to-end trainable CNN based on a re-formulated atmospheric scattering model.
  • C. Main Contributions: AOD-Net can be embedded with other deep models and jointly tuned end to end for high-level tasks on hazy images.
  • C. Main Contributions: AOD-Net processes one 480 × 640 image in as little as 0.026 second on a single GPU.
  • C. Main Contributions: The paper extends an earlier conference version with deeper object-detection analysis, joint training details, architecture analysis, and broader comparisons.

II. AOD-NET: THE ALL-IN-ONE DEHAZING MODEL

AOD-Net reformulates the atmospheric scattering model so one input-adaptive variable replaces separate transmission and atmospheric-light estimation, enabling direct clean-image reconstruction.

  • A. Physical Model and Transformed Formula: The atmospheric scattering model represents a hazy image as scene radiance attenuated by transmission plus atmospheric light.
  • A. Physical Model and Transformed Formula: The conventional pipeline separately estimates transmission, estimates atmospheric light, and then reconstructs the clean image, allowing errors to accumulate or amplify.
  • A. Physical Model and Transformed Formula: AOD-Net unifies transmission and atmospheric light into K(x) and directly minimizes pixel-domain reconstruction errors.
  • A. Physical Model and Transformed Formula: K(x) integrates 1/t(x) and A, while b is a constant bias with default value 1.
  • A. Physical Model and Transformed Formula: The network estimates K(x) from the hazy input and uses it as input-adaptive parameters to estimate the clean image.

B. Network Design

AOD-Net combines K estimation with clean-image generation in a lightweight, multi-scale CNN. Jointly estimating transmission and atmospheric light avoids overexposure and supports more realistic restoration.

  • Network Design: AOD-Net uses a K-estimation module followed by clean-image generation with input-adaptive parameters.The first module estimates K(x) from I(x), and the second estimates J(x) using K(x).
  • Network Design: The K-estimation module fuses varied-size-filter features through inter-layer concatenations to capture multiple scales and compensate for convolutional information loss.Each convolutional layer uses only three filters, making the network lightweight relative to existing deep methods.
  • Network Design: Jointly estimating t(x) and A enables mutual refinement, producing more realistic lighting and structural details than a baseline that estimates A separately.The separate-estimation baseline overestimates A and causes overexposure.

A. Datasets and Implementation

The evaluation uses synthesized hazy images with known ground truth, standard training settings, and comparisons against established dehazing methods. Results are assessed with PSNR, SSIM, visual quality, and error decomposition.

  • Datasets: Synthetic haze is generated from NYU2 depth images using atmospheric lights in [0.6, 1.0] and β values from 0.4 to 1.6.The data include 27,256 training images, 3,170 NYU2 TestSet A images, and 800 Middlebury TestSet B images.
  • Implementation: Training uses ReLU activations, batch size 8, learning rate 0.001, momentum 0.9, decay 0.0001, and MSE loss.The model is trained for 40 epochs, although it usually converges after around 10 epochs.
  • Evaluation: Comparisons include FVR, DCP, BCCR, ATM, CAP, NLD, DehazeNet, and MSCNN, with PSNR and SSIM available for synthesized images because ground truth is known.Real hazy images generally lack haze-free ground truth for quantitative restoration evaluation.
  • Quantitative Results: AOD-Net achieves higher PSNR and greater SSIM advantages than competitors on TestSets A and B.The authors attribute the PSNR advantage partly to MSE optimization and investigate the stronger SSIM improvement separately.
  • Error Analysis: AOD-Net has similar residual MSE to DehazeNet and CAP but substantially lower mean-image MSE, indicating better recovery of global illumination.The mean component relates to global illumination, while the residual reflects local structural variations and contrasts.
  • SSIM Analysis: AOD-Net’s mean contrast value is 0.9989 on five TestSet B images, exceeding the reported values for all listed competing methods.The comparison includes ATM at 0.7281, BCCR at 0.9574, FVR at 0.9630, and CAP at 0.9760, among others.

C. Qualitative Visual Results

AOD-Net produces visually faithful results across synthetic, challenging natural, and white-scene images, while also preserving haze-free inputs and showing promise for anti-halation enhancement.

  • Synthetic Images: On synthetic images, AOD-Net generally yields sharper contours, richer colors, and greater visual faithfulness to ground truth.Figure 5 compares hazy inputs, DehazeNet, MSCNN, AOD-Net, and ground-truth images.
  • Challenging Natural Images: On challenging natural images, AOD-Net generalizes from indoor synthetic training to cluttered outdoor scenes with fine textures and illumination variations.The cited comparisons report artifacts from FVR and unrealistic color tones from several other methods.
  • White Scenery Natural Images: AOD-Net is almost artifact-free on white scenes, avoiding the fake colors and distorted contours seen in competing results.DCP shows especially strong sky artifacts, while CAP, DehazeNet, and MSCNN retain or introduce other visible problems.
  • Haze-Free Images: AOD-Net causes little negative impact when applied to haze-free images, supporting the reported robustness of its K-estimation module.Figure 8 shows outputs on two challenging clean images.
  • Image Anti-Halation: Without retraining, AOD-Net produces decent anti-halation enhancement results on photos with spreading light in bright regions.Figure 9 compares real halated photos with AOD-Net outputs.

D. Effectiveness of Multi-Scale Features

This section examines inter-layer concatenations in the K-estimation module and compares them with a sequential baseline. The supplied evidence identifies the evaluation setup but does not provide the complete comparison outcome.

  • Effectiveness of Multi-Scale Features: The study tests whether inter-layer concatenations of multi-scale features improve K-estimation compared with a sequential five-convolution baseline.The baseline is “conv1 →conv2 →conv3 →conv4 →conv5(K)” without inter-layer concatenation.
  • Effectiveness of Multi-Scale Features: The supplied passage reports the baseline’s TestSet A average PSNR as 17.0517 dB, while the corresponding SSIM statement is truncated.No complete AOD-Net comparison value is available in the supplied passage.
  • Additional Evaluations: The supplied evidence includes figures for white-scene dehazing, haze-free inputs, anti-halation, and model running time, but does not state their comparison results.The captions identify the visual content or metric scope only.

E. Running Time Comparison

The paper evaluates haze-robust object detection on synthetic Pascal VOC data and reports gains from dehazing, joint tuning, and a unified model across haze levels.

  • Quantitative detection results: AOD-Net+Faster R-CNN performs comparably to MSCNN+Faster R-CNN without joint tuning but slightly worse than DCP+Faster R-CNN.
  • Joint optimization: mAP increases from 0.5794 to 0.6819 for heavy haze after jointly tuning AOD-Net+Faster R-CNN end to end.
  • Joint optimization: Retrained Faster R-CNN rises from 0.5155 to 0.6756 under heavy haze but remains consistently worse than JAOD-Faster R-CNN.
  • Multiple haze levels: A model trained with haze levels sampled using β from [0, 0.1] performs consistently across three haze levels, though slightly below dedicated models.

B. Visualized Results on Natural Hazy Images

Natural-image visualizations compare six detection pipelines and show that jointly tuned AOD-Net improves detection beyond standalone dehazing or non-joint alternatives.

  • Compared methods: Six approaches are compared, including naive, dehazing-augmented, fine-tuned, and jointly tuned Faster R-CNN pipelines.
  • Observed effects of haze: Haze causes missing detections, inaccurate localizations, and unconfident category recognitions for Faster R-CNN.
  • Visual comparison: AOD-Net+Faster R-CNN shows visible advantages over naive Faster R-CNN, while JAOD-Faster R-CNN improves further and visibly surpasses the alternatives.The figure uses a confidence threshold of 0.6; fine-tuned models use synthetic light haze training.
  • Joint optimization: The pipeline's advantage comes from jointly estimating dehazing parameters and tuning dehazing with detection and recognition end to end.

or just adding more parameters?

The paper tests whether JAOD-Faster R-CNN's gains arise merely from extra parameters by replacing AOD-Net with an equally sized plain auto-encoder.

  • Controlled baseline: The comparison is designed to test whether adding layers and parameters without task-specific dehazing improves object detection under haze.
  • Controlled baseline: Auto-Faster R-CNN replaces AOD-Net with a plain convolutional auto-encoder having exactly the same parameter count.The auto-encoder uses five convolutional layers and resembles the K-estimation module's structure.
  • Efficiency: JAOD-Faster R-CNN adds little runtime over retrained Faster R-CNN: 0.192s versus 0.166s per image on a GTX TITAN X GPU.
  • Conclusion: The paper concludes that AOD-Net combines superiority, robustness, and efficiency while supporting jointly tuned detection on hazy images.
Loading 1707.06543v1…