Benchmarking Single Image Dehazing and Beyond (1712.04143v4)

Abstract: We present a comprehensive study and evaluation of existing single image dehazing algorithms, using a new large-scale benchmark consisting of both synthetic and real-world hazy images, called REalistic Single Image DEhazing (RESIDE). RESIDE highlights diverse data sources and image contents, and is divided into five subsets, each serving different training or evaluation purposes. We further provide a rich variety of criteria for dehazing algorithm evaluation, ranging from full-reference metrics, to no-reference metrics, to subjective evaluation and the novel task-driven evaluation. Experiments on RESIDE shed light on the comparisons and limitations of state-of-the-art dehazing algorithms, and suggest promising future directions.

• The paper introduces the RESIDE dataset to benchmark dehazing using both synthetic and real-world images.
• It systematically evaluates nine state-of-the-art algorithms with full-reference, no-reference, and task-driven metrics.
• Findings reveal that while CNN-based methods excel in objective metrics, classical approaches often yield superior subjective quality.

Benchmarking Single Image Dehazing and Beyond

The paper "Benchmarking Single Image Dehazing and Beyond" presents a thorough evaluation of contemporary single image dehazing algorithms, facilitated by the introduction of a novel large-scale benchmark dataset named REalistic Single Image DEhazing (RESIDE). This dataset is meticulously curated to include diverse image contents and data sources, enabling a comprehensive assessment through various evaluation criteria that span full-reference metrics, no-reference metrics, subjective evaluations, and task-driven evaluations.

Problem Description

Single image dehazing addresses the prevalent issue of poor visibility in outdoor images caused by aerosol-induced haze. The degradation effects from haze, including reduced contrast and color shift, challenge the accuracy and effectiveness of subsequent computer vision tasks, such as object detection in autonomous driving scenarios. Consequently, single image dehazing has garnered substantial attention within image restoration and enhancement domains.

The atmospheric scattering model serves as the fundamental framework for dehazing, described by the equation:

$I(x) = J(x)t(x) + A(1 - t(x)),$

where $I(x)$ represents the observed hazy image, $J(x)$ is the haze-free scene radiance, $A$ denotes global atmospheric light, and $t(x)$ is the transmission matrix, which is further defined as:

$t(x) = e^{-\beta d(x)},$

with $\beta$ being the scattering coefficient and $d(x)$ the depth between object and camera. This model guides the estimation of the clean image $J(x)$:

$J(x) = \frac{1}{t(x)} I(x) - A \frac{1}{t(x)} + A.$

Existing Methodology

Contemporary dehazing methods predominantly adhere to a three-step process: estimating the transmission matrix $t(x)$, estimating $A$, and computing $J(x)$. Both physically grounded priors and data-driven approaches facilitate the estimation of these parameters. Several earlier methods leveraged natural image priors, such as the Dark Channel Prior (DCP), color attenuation prior, and non-local prior assumptions.

The late integration of Convolutional Neural Networks (CNNs) represents a significant evolution in dehazing methodologies. Models such as DehazeNet and MSCNN use CNNs to directly infer $t(x)$ from hazy images, enhancing performance by bypassing the inaccuracies inherent in manual parameter estimation. Moreover, fully end-to-end models like AOD-Net further streamline dehazing by directly outputting the clean image without intermediary estimations.

Contribution and Benchmarking

This paper provides several significant contributions:

1. Introduction of RESIDE Dataset:
   • RESIDE incorporates 13,990 synthetic training images, alongside 500 synthetic and 20 hybrid (synthetic and real-world) testing images. The RESIDE-$\beta$ subset further includes 72,135 outdoor training images and a task-driven evaluation set (RTTS) with 4,322 annotated real-world hazy images.
2. Diverse Evaluation Criteria:
   • Evaluations utilize not only traditional full-reference metrics (PSNR, SSIM) but also no-reference metrics (SSEQ, BLIINDS-II) and subjective assessments. Additionally, the paper introduces task-driven evaluations to bridge the gap between dehazing quality and practical utility.
3. Systematic Experimental Analysis:
   • Nine state-of-the-art dehazing algorithms are evaluated across various metrics. The paper found that while CNN-based methods like DehazeNet and AOD-Net lead in PSNR and SSIM, classical prior-based methods often perform better in subjective quality assessments. Moreover, DehazeNet and CAP exhibited the lowest perceptual loss, aligning better with high-level visual tasks.

Implications and Future Directions

The results underscore the nuanced effectiveness of different dehazing algorithms, revealing inconsistencies between traditional metrics and practical performance. For instance, the multi-scale design of MSCNN is favored in both human perception and high-level vision tasks. Moreover, the paper signals the need for real-world generalization and evaluation methods that align closely with practical applications in adverse conditions like haze.

Future research should pivot towards:

1. Developing no-reference metrics that correlate better with human perception.
2. Optimizing dehazing algorithms with dedicated high-level vision tasks in mind.
3. Enhancing the realism of synthetic training datasets, or leveraging domain adaptation techniques to align synthetic and real-world data.

The practical applications of this research span various domains where visibility is crucial, from autonomous vehicles to surveillance systems, signifying its wide-reaching impacts and the growing importance of robust dehazing solutions.