Single hazy image restoration using robust atmospheric scattering model
Introduction
Outdoor images captured in adverse weather are attenuated due to atmospheric scattering and absorption during the propagation of light. The scattering is caused by water droplets or suspended particles in the atmosphere, which veils the details of images [1]. The absorption is an attenuation phenomenon of light, which diminishes the visibility of images [2]. Consequently, images taken in poor weather usually present degradation both in contrast and color.
Nowadays, many crucial applications of computer vision, such as object detection and automatic driving, rely heavily on the quality of outdoor image. Thus, researches on the approach of haze removal (dehazing) have become a subject of great significance and interest. However, dehazing just using a single hazy image is an under-constrained problem, for the thickness of haze in outdoor scenes is dependent on unknown atmospheric light and scenes depth. Accordingly, several methods based on multiple images or dedicated hardware have been proposed for image dehazing. Schechner et al. [3] employed a polarization camera to capture multiple images with different polarization angles at the same scene, and calculated the atmospheric light and scene depth from these images to restore clear image. Liang et al. [4] proposed a polarimetric dehazing algorithm which fuses infrared and visible images to improve the visual quality of hazy image.
In recent years, great advances have been achieved in single image dehazing, for quite a few valid priors or assumptions were exploited. Tan [5] has found that haze-free image with fine visual effect presents relatively high contrast, by which he successfully enhanced the contrast of hazy image. Galdran [6] innovatively fused a series of images which were obtained by artificially reducing exposure of a single input to achieve dehazed image. Galdran et al. [7,8] also employed the Total Variational technique to enhance the contrast of hazy image. Nishino et al. [9] introduced a Bayesian Probabilistic algorithm which could jointly estimate the depth and scene albedo from a single image. Moreover, an image fusion scheme [10,11] which fuses two enhanced versions of an original image has been introduced by several researchers to perform contrast enhancement. Conclusions can be draw from plenty of practical applications that most of the image enhancement methods are uncomplicated to implement, for these methods do not depend on atmospheric scattering model. Nevertheless, when such algorithms are applied to process hazy image, a truth is neglected that the degradation of images is in proportion to distance from objects to the camera.
On the other hand, numerous physics-based methods were proposed as well, which are aimed at inversely solving the optical model to recover degraded image. Among these approaches, dark channel prior (DCP) [12] that some pixels with low intensities exist at least one channel in RGB space is considered as the most famous one. With this prior, the thickness of haze can be estimated accurately, and haze-free image can be obtained by utilizing the atmospheric scattering model. On this foundation, Peng et al. [13], Nair and Sankaran [14] and Huang et al. [15] ameliorated DCP to achieve more excellent effectiveness in haze removal. Besides, Li and Zheng [16] attempted to restore hazy image by exploiting Globally Guided Image Filtering (G-GIF). Berman [17,18] proposed the non-local prior (NLP) which describes the change of pixels values from scene radiance to observed image to estimate the transmission map and atmospheric light. Meng et al. [19] introduced a novel boundary constraint and contextual regularization to calculate the transmission map (BCCR). Fattal [20] discovered a generic distribution regularity of small local patches of hazy image, known as color-lines, and applied it to derive the scene transmission from a single original image. Zhu et al. [21] calculated the weighted sum of saturation, brightness and error as the scene depth, called the color attenuation prior. Ki et al. [22] minimized the information loss of output image to obtain the transmission map with high quality. To achieve expected local contrast gain, Ling et al. [23] developed a novel transmission model to flexibly adjust the level of haze removal. In addition, noise filtering [24], regression model [25] and image segmentation have been used for dehazing. For instance, Zhu and He [26] accurately estimated the transmission map by utilizing graph cut algorithm (GC).
Recently, convolutional neural network (CNN) has witnessed remarkable achievements in image classification tasks and was applied to haze image restoration. Ren et al. [27,28] proposed an effective multi-scale CNN (MSCNN) and a gated fusion network (GFN) to restore high quality hazy-free images. Ren et al. [29] also realized single video dehazing by combining CNN with semantic segmentation. Li et al. [30] exploited an end-to-end dehazing CNN, called AOD-Net, which can optimize the end-to-end pipeline from inputs to outputs for the first time. Zhang and Patel [31] proposed densely connected pyramid dehazing network (DCPDN) which can learn atmospheric light, transmission map and dehazing simultaneously.
Although the above methods achieve a good performance in terms of haze removal, these methods share a common limitation that the global brightness of restored images is relatively low, especially for images with uneven illumination. To overcome this problem, a robust atmospheric scattering model (RASM) is proposed for hazy image restoration in this paper. We innovatively decompose the scene radiance into reflectance and incident light and introduce a noise term in traditional atmospheric scattering model (TASM). Moreover, in order to alleviate the impact of over-enhancement in dense haze regions, a simple and effective compensation term is added to the transmission map. Then an alternating iterative method is applied to simultaneously calculate the incident light and reflectance. Due to reasonable assumptions with respect to RASM, the restored images are characterized by high contrast, clear visibility and natural appearance. Furthermore, multitudes of hazy images are employed to demonstrate our approach has a competitive performance for different scenes.
We organize the remainder of this paper as follows. In Section 2, RASM is analyzed and presented. In Section 3, inputs of RASM (the transmission map and atmospheric light) are calculated. In Section 4, our image restoration algorithm is introduced in details. In Section 5, experimental results are illustrated. In Section 6, the conclusion is narrated.
Section snippets
Traditional atmospheric scattering model
According to Refs. [32], [33], [34], the traditional atmospheric scattering model widely used for processing hazy images can be expressed as:where x represents the pixel coordinate. c represents the color channels in the RGB space. Lc(x) is the observed image. S c(x) is the clear image or scene radiance. B c denotes the atmospheric light. t(x) denotes the transmission map which is distance-dependent. t(x) can be given by:where d(x) is the depth
Estimation of the atmospheric light and the transmission map
An accurate estimation of the atmospheric light and the transmission map is pivotal for image restoration. In this section, estimation algorithms of the atmospheric light and the transmission map are presented in details.
Restoring image by solving objective function
In this section, a method to restore the haze-free image based on the RASM will be introduced in details. When the transmission map and the atmospheric light are derived, they can be used as inputs of TASM to directly restore a clear image. However, for RASM, it is unfeasible to calculate three unknowns with just two inputs (transmission map and atmospheric light). Therefore, we formulate an objective function by taking some prior information into consideration and adopt the alternating
Experiment
In this section, the performance of the proposed method was validated by experimental results with the implementation of qualitative comparison, quantitative evaluation, subjective evaluation and algorithm complexity analysis. Several existing state-of-the-art haze removal methods were employed to compare with the proposed method, such as DCP [12], NLP [17], BCCR [19], GC [26], MSCNN [27] and AOD-Net [30]. Three well-known data-bases, REalistic Single Image DEhazing (RESIDE) dataset [40],
Conclusion
The traditional atmospheric scattering model used for image dehazing may fail to improve brightness of the dehazed image. To address this problem, we propose a robust atmospheric scattering model by decomposing the real scene into the incident light and the reflectance and attaching an additional term to the traditional atmospheric scattering model. Then the objective function is formulated based on the proposed robust atmospheric model, and some novel regularization terms are imposed in our
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
This work is supported by Key Research and Development Projects of Shandong Province (Industry Key Technology) (Grant No. 2016CYJS02A01) and Shandong Key Research and Development Program (public welfare special) (Grant No. 2017GGX30103).
References (49)
- et al.
Haze removal based on multiple scattering model with superpixel algorithm
Signal Process.
(2016) Image dehazing by artificial multiple-exposure image fusion
Signal Process.
(2018)- et al.
A multi-scale fusion scheme based on haze-relevant features for single image dehazing
Neurocomputing
(2018) - et al.
Color image dehazing using surround filter and dark channel prior
J. Vis. Commun. Image Represent
(2018) - et al.
Improved algorithm for image haze removal based on dark channel priority
Comput. Electr. Eng.
(2018) - et al.
Perception oriented transmission estimation for high quality image dehazing
Neurocomputing
(2017) - et al.
Image de-hazing from the perspective of noise filtering
Comput. Electr. Eng.
(2017) - et al.
Fast single image dehazing based on a regression model
Neurocomputing
(2017) - et al.
A fast image dehazing algorithm based on negative correction
Signal Process.
(2014) - et al.
Haze removal method for natural restoration of images with sky
Neurocomputing
(2018)
Single image haze removal based on the improved atmospheric scattering model
Neurocomputing
A fusion-based enhancing method for weakly illuminated images
Signal Process.
A spatial processor model for object color perception
J. Frankl. Inst.
No-reference image quality assessment based on spatial and spectral entropies
Signal Process.: Image Commun.
Edge-preserving decomposition-based single image haze removal
IEEE Trans. Image Process.
Polarization-based vision through haze
Appl. Opt.
Polarimetric dehazing method for visibility improvement based on visible and infrared image fusion
Appl. Opt.
Visibility in bad weather from a single image
Enhanced variational image dehazing
SIAM J. Imaging Sci.
Fusion-Based variational image dehazing
IEEE Signal Process. Lett.
Bayesian defogging
Int. J. Comput. Vis.
Single image dehazing by multi-scale fusion
IEEE Trans. Image Process.
Single image haze removal using dark channel prior
Generalization of the dark channel prior for single image restoration
IEEE Trans. Image Process.
Cited by (23)
Recent advances in image dehazing: Formal analysis to automated approaches
2024, Information FusionIncorporating inconsistent auxiliary images in haze removal of very high resolution images
2023, International Journal of Applied Earth Observation and GeoinformationSingle low-light image enhancer using Taylor expansion and fully dynamic convolution
2021, Signal ProcessingSingle underwater image restoration by decomposing curves of attenuating color
2020, Optics and Laser TechnologyA comprehensive survey on image dehazing for different atmospheric scattering models
2024, Multimedia Tools and Applications