1. Introduction
Photographs taken in natural settings frequently encounter adverse conditions. Weather fluctuations, atmospheric contaminants, haze, fire, smoke, and dust can degrade image quality. Hazy images suffer from reduced visibility due to fog, dust scattering, and light attenuation, complicating object identification and color fidelity. Addressing the challenges posed by hazy images constitutes a significant area of research in image processing, with various approaches being explored [
1,
2,
3,
4,
5,
6,
7,
8,
9,
10,
11,
12]. Techniques for haze removal, such as Markov random field (MRF) [
1], filtering [
3,
5], color analysis [
2,
7,
9,
10], transmission map estimation [
4,
6,
8], and deep learning approaches [
11,
12], play pivotal roles in enhancing visibility compromised by haze in the resulting images.
He et al. [
4] presented a seminal work on transmission map estimation using the dark channel prior (DCP) notion. This method capitalizes on the observation that, within local regions of outdoor images, at least one color channel typically exhibits markedly low intensity, which they term the
dark channel. They noted that the dark channel tends to become brighter in the presence of haze. The algorithm effectively estimates and mitigates fog, leveraging this characteristic alteration and enhancing visibility and image clarity in outdoor settings.
However, in situations where the color of the target object closely matches atmospheric light and lacks shadows, the DCP exhibits constraints, potentially misidentifying such instances as haze. Additionally, during haze depth estimation, the transmission map may incorrectly interpret sky regions as heavily fogged, leading to inaccuracies. The restoration process could introduce image distortions in areas with significant fog due to over-correction. Moreover, utilizing a patch size of for DCP computation might leave halo artifacts.
Various efforts have been made to enhance the DCP-based transmission map estimation approach [
6,
8,
13]. These refinements are geared towards enhancing the precision of transmission map estimation. While techniques such as boundary constraints [
6], multiscale Laplacian and Gaussian pyramids [
8], and dual transmission map estimation [
13] show promise, challenges persist in addressing night-time haze and mitigating halo artifacts.
Drawing from the insights of [
4], we extend the discourse with fresh perspectives and innovations by proposing an approach for image dehazing. We introduce a method for estimating enhanced transmission maps utilizing multiple
DCPs derived from multiscale hazy, inverted, and Euclidean difference images which strive to enhance the granularity of transmission maps. To tackle the challenge posed by low-light, hazy images, we propose a new technique for airlight estimation. Furthermore, we introduce an adaptive gamma correction method to refine the transmission map further.
Our proposed method outperforms the next-best technique by achieving a significantly higher score of approximately 3.03 points on the Dehazing Quality Index (DHQI) [
14] within a Real-World Task-Driven Testing Set (RTTS) [
15].
3. Preliminary
He’s image dehazing method [
4] revolves around generating a DCP in a hazy image, estimating the transmission map, and then recovering it to produce a dehazed image. The DCP is computed by selecting the minimum value among the RGB channels for each pixel and further refining it locally using a patch size of
. The brightest pixel within this dark channel image is the basis for estimating the airlight present in the scene. The transmission map
, indicative of the ratio of light reaching the observer from objects within the scene, is estimated as follows:
where
is a preservation constant,
is the hazy image of a color channel
c,
is the estimated airlight of
c, and
is the patch centered on pixel
x.
4. Proposed Method
Despite the widespread adoption of the DCP-based transmission map estimation approach for image dehazing [
4] and its various extensions [
6,
8,
13], they still exhibit notable limitations in image quality. This paper introduces an image dehazing method inspired by [
4], as outlined in Algorithm 1. Our proposed method harnesses multiscale and multiple features to compute the
DCPs for better transmission map estimation. Additionally, we incorporate an enhanced airlight estimation method and adaptive gamma correction to refine the estimation of the transmission map, aiming to address the quality issues of the image above.
Algorithm 1 Enhanced Transmission Map with Multiscale Multiple Features DCP and Adaptive Gamma Correction |
Input: —Hazy Image Output: —Dehazed Image Procedure DehazingProcess() Step 1: Resize to generate and of Step 2: Estimate an airlight adaptively using Equation ( 6). Multiscale multiple features transmission map estimation ▹ Begin Multi-scale multiple features transmission map estimation Step 3: Generate three DCPs based on , and of . Step 4: Estimate three intermediate transmission maps based on DCPs and airlight from steps 2 and 3, followed by a Guided filtering. Step 5: Each intermediate transmission map is rescaled to match the original size before being averaged, resulting in a transmission map . Step 6: Generate inverted and Euclidean difference images with Equations ( 2) and ( 3). Step 7: Repeat step 3 to step 5 to yield , and . ◃ End Multi-scale multiple features transmission map estimation Step 8: Utilizing , , , and the image derived from Equation ( 4), an enhanced transmission map is computed using Equations ( 7)–( 9). Step 9: Refine by adaptive gamma correction using Equation ( 10). Step 10: Recover the final image, based on in step 9.
|
4.1. 1 × 1 Dark Channel Priors
This section delineates the derivation of the multiple DCPs from multiscale hazy, inverted, and Euclidean difference images that constitute the core component of our proposed method.
4.1.1. Inverted and Euclidean Difference Images
Hazy images often display notable differentiation between low-light and bright regions, resulting in significant discrepancies during transmission map estimation. To address this issue, we propose employing image inversion to enhance low-light portions and darken bright ones, thereby alleviating distortions caused by haze. The image inversion is carried out channel-wise and defined as follows:
where
is the inverted image.
To measure the disparity between the original and inverted images, we leverage the Euclidean difference image
, which is defined as follows:
where
is a grayscale image, and
is the stretched
.
The Euclidean difference image improves the visibility of haze-affected regions and enhances transmission map estimation precision.
For a unified analysis, the Euclidean difference images from R, G, and B channels are combined as follows:
where
is the aggregate Euclidean difference image, enhancing the estimation’s accuracy and reducing distortions in the haze removal process.
4.1.2. Intermediate Transmission Maps
In He’s approach [
4], DCP derived from the
patch size led to residual haze persisting around objects (halo artifacts). Conversely, DCP with
patch size focuses solely on pixel values, disregarding object boundaries and local patterns due to the absence of neighboring pixel information. Additionally, guided filters [
18] coupled with multiscale processing are introduced to estimate a more precise transmission map.
Specifically, as depicted in
Figure 1, a hazy image
is first resized to
and
of its original size. Subsequently, three distinct dark channels with a patch size of
are computed from these two resized images and
. Intermediate transmission maps for each dark channel are estimated with (
1), along with our proposed airlight estimation method (
Section 4.2). A guided filter [
18] is then applied on each intermediate transmission map. Following this, the scaled intermediate transmission maps are restored to match the dimensions of
and aggregated through averaging to yield
.
The same procedure is iterated for inverted image
and channel-wise Euclidean difference images
, resulting in intermediate transmission maps
and
, respectively. Multiscale processing and averaging are also applied to the
obtained in (
4).
4.2. Airlight Estimation
In [
4], airlight estimation entails selecting the brightest top 1000 pixels, a method pivotal for distinguishing foreground from background in hazy images based on brightness. However, applying this technique to inherently low-light images can lead to excessive darkening of the images, thereby compromising the retention of original features and risking the loss of image details.
To resolve this challenge, we leverage the brightest and the darkest 1% of pixels to estimate airlight. The process can be outlined as follows: First, calculate the average brightness of
, denoted as
, and normalize it to the range of
through division by 255. Secondly, determine the number of pixels, denoted as
S, that corresponds to 1% of the total number of pixels in the image. Lastly, estimate the number of dark pixels as follows:
where
is the dark pixel count, and the bright pixel count is obtained by subtracting
from
S. We select the darkest
pixels with
and the brightest
pixels through
from the dark channel image.
The estimated airlight
can be determined from the following:
where
represents the color of pixel
i, and
and
are the indices of the darkest and brightest pixels, respectively.
4.3. Enhanced Transmission Map
For image dehazing, precise identification of haze presence and intensity is paramount for enhancing image quality. By taking a product of
and
, we aim to preserve authentic foreground features while enhancing image contrast, as depicted in
Figure 2a. However, the product operation may lead to a darkening effect, necessitating a compensatory stretching process. Thus,
is determined as follows (
Figure 2b):
Nonetheless, such contrast adjustment could potentially obscure or eliminate specific image details. Therefore, we take an average of
and
, yielding
, which results in darker structural elements and brighter hazy regions, as illustrated in
Figure 2c.
While this operation aligns with hazy image characteristics, it comes at the expense of reduced structural intricacy. To address this issue, an enhanced transmission map,
, as shown in
Figure 2d, is obtained by averaging the
with
:
The averaging process enhances structural details and mitigates distortions in the sky region. Structures farther away are typically shrouded in denser fog, necessitating more extensive restoration efforts. Conversely, corrections applied to the sky region are relatively modest. Meanwhile, structures in closer proximity are presumed to encounter lighter fog, thus undergoing a more subdued correction process.
4.4. Adaptive Gamma Correction
Low-valued transmission maps may lead to distortions in dehazed images, potentially causing darkening. For example, in
Figure 2d, shallow pixel values indicate a risk of the dehazed image appearing similarly shallow and darker if not corrected. Adjusting the transmission map is necessary to address this issue and achieve proper brightness correction.
Gamma correction is a remedy that involves adjusting pixel values in an image to compensate for the non-linear way human eyes perceive light through a
parameter [
19]. As shown in
Figure 3 and
Figure 4, when
is increased, images become darker, gradually reducing fog, whereas decreasing
retains more fog. An improper selection of
may result in insufficient darkening or ineffective fog removal. Therefore, choosing an appropriate
should consider the characteristics of the fog image. Instead of manual adjustment, we propose an adaptive approach to determine the
value based on the maximum and minimum values of the enhanced transmission map
:
To prevent the maximum and minimum values from being saturated, is set to 0.1.
4.5. Image Dehazing
Throughout the haze removal process, the hazy image
undergoes traversal across its color dimensions alongside the estimated airlight
and
. A dehazed image
can be obtained via the following:
where
is the threshold to preserve the transmission map and is set to 0.1 in this paper. The dehazing process mitigates the influence of airlight from each channel, restoring the lost color and brightness of the hazy image.
5. Evaluations
The proposed method is assessed using the Real-World Task-Driven Testing Set (RTTS) obtained from the Realistic Single Image Dehazing (RESIDE) benchmark [
15]. The RTTS is specifically curated to consist solely of images depicting natural haze, ensuring the authenticity and relevance of the evaluation. This dataset comprises 4322 images characterized by diverse dimensions and a wide range of haze intensities and environmental conditions.
In our evaluation, we utilize the Dehazing Quality Index (DHQI) [
14], a metric designed to assess the effectiveness of dehazing algorithms specifically on images affected by fog or hazy weather, rather than evaluating overall noise reduction.
5.1. Comparisons with Existing Works
As depicted in
Figure 5, our method effectively removes haze from the image while preserving the integrity of the sky region. Furthermore, as demonstrated in
Table 1, our proposed approach outperforms the second-highest scorer, the method of Tarel and Hautiere [
3], by approximately 3.03 in DHQI.
A comparison between our method and that of He et al. [
4] reveals significant disparities. He et al.’s approach leads to color distortion within the sky region and excessive removal of haze around objects, characterized by visible white bands, as depicted in
Figure 6. In contrast, our method exhibits no such distortions in the sky and avoids the formation of band-like haze around objects, as detailed in
Table 1.
While deep-learning-based approaches have shown success in various applications, including image dehazing [
11] and restoration [
12], they do not demonstrate a clear advantage in realistic hazy image datasets such as the RTTS and image dehazing metric DHQI compared to our method. Deep-learning-based methods often rely on the training dataset, leading to a performance influenced by the data-driven approach.
In contrast, our proposed method directly performs adaptive dehazing from the original image, yielding consistent results under various conditions. This method also possesses the capability to analyze the image’s state in real time and execute optimal dehazing without depending on specific training data.
5.2. Ablation Study
In this section, we conduct an ablation study to assess the impact of various components introduced in our method on dehazing performance.
Table 2 shows that the proposed enhanced transmission map generated from inverted and Euclidean difference images contributes most significantly, outperforming He’s approach [
4] by 7.59 points despite the absence of guided filtering, adaptive AR estimation, multiscale processing, or adaptive gamma correction.
In
Figure 7c, residual haze can be observed surrounding objects, suggesting additional refinement is necessary. We employed a
DCP alongside a guided filter to tackle this issue. While this approach led to a marginal decrease of approximately three points in the DHQI, as shown in
Table 2, the integration of adaptive gamma correction yielded an approximate one point improvement. Moreover, incorporating multiscale processing techniques and adaptive AR estimation contributed to further performance enhancement, as illustrated in
Figure 7d.
Figure 8 visually demonstrates the efficacy of employing a
DCP and guided filter combination for more precise and detailed dehazing, consequently improving object detection performance. The contrast between using He et al. [
4] and our
DCP and guided filter is highlighted in
Figure 8b and
Figure 8c, respectively. Additionally, as depicted in
Figure 8d, a multiscale approach boosts object detection confidence.
Ultimately, integrating all components, our method achieves a notable numerical improvement of approximately nine points over the approach by He et al. [
4].
5.3. Discussion
Our approach excels in images with light-to-moderate haze, ensuring that object features remain clearly visible. However, as demonstrated in row #4 of
Figure 9l, when haze becomes too dense, there is a heightened risk of losing structural details, leading to incomplete haze removal. This limitation is particularly noticeable with distant structures heavily obscured by thick haze, where the fine details become challenging to recover, making it difficult to remove the haze effectively while preserving the clarity of objects. Furthermore, as shown in row #7 of
Figure 9l, using a guided filter can occasionally result in a stair-step effect in sky regions with sharp color transitions.
Thus, the most suitable images for this method are those with moderate haze, where the objects remain distinguishable. Achieving effective dehazing without compromising object visibility becomes increasingly challenging when the haze becomes too thick.
6. Conclusions
This paper introduces a novel method for image dehazing employing the DCP-based transmission map estimation technique. Our approach incorporates several innovative elements to derive an enhanced transmission map, such as employing multiple DCPs derived from multiscale hazy, inverted, Euclidean difference images, guided filtering, airlight estimation, and adaptive gamma correction. The proposed method outperformed He’s method by 9 points and achieved a 3.03 point higher DHQI score than the second-best approach. Additionally, it significantly reduced distortion in the sky region, surpassing all existing state-of-the-art techniques. However, effectively eliminating highly dense haze remains a persistent challenge, highlighting the need to explore further fog removal techniques tailored to diverse environmental conditions.