Single hazy image restoration using robust atmospheric scattering model

Highlights

• The robust atmospheric scattering model is proposed to restore hazy images.

• A compensation term attached to transmission map can suppress halo artifacts.

• An objective function with novel regularization terms is formulated.

• The illumination of restored images can be significantly improved.

Abstract

Images captured in unfavorable weather usually exhibit poor visibility, which results from scattering and absorption that the propagated light suffers in the atmosphere. To improve the quality of degraded images, multitudes of algorithms have been exploited based on traditional atmospheric scattering model. However, in the traditional model, a phenomenon is neglected that the radiance projected on scenes is uneven, which leads to low brightness in processed image. Targeted the inherent limitation of the traditional model, we propose a robust atmospheric scattering model by decomposing the real scene into incident light and reflectance component and attaching a noise term in the traditional model. Then an objective function which includes novel regularization terms for the incident light and reflectance is formulated based on the proposed model, and an alternating direction method of multipliers is adopted to jointly estimate the incident light and reflectance. Moreover, a compensation term with regard to transmission map is introduced to ameliorate over-enhancement in thick haze regions. Ultimately, comprehensive tests are implemented to compare our method with other exceptional haze removal methods. Experiments on images with different characteristics manifest excellent performance of the proposed method in terms of haze removal and brightness enhancement.

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.