Effective haze removal under mixed domain and retract neighborhood

Abstract

The quality of images would exhibit degraded visibility during inclement weather conditions. We proposed a novel method for estimating an optimal transmission map and recovering the real scene. Under HSI color model, saturation layer and intensity layer are mixed together for obtaining the rough transmission. To avoid halos and artifacts, proposed approach employs edge preserving constraint of shrinkage neighborhood on the color line model, which can maintain maximum smoothness and sharp edges in the refined transmission map. A comparative experiment with a few previous methods shows improvement visual results.

Introduction

The quality of images would exhibit degraded visibility during inclement weather conditions. Therefore, outdoor scenes usually have low visibility and contrast, which lead to depressed reliability of many vision systems such as the automatic vehicle driving system and the intelligent navigation system. Haze removal just decreases the influence of the haze and tries to restore such this image by recovering the aesthetic quality and also improving the details (Fig. 1).

Normally, numerous dehazing approaches are based on physical degradation model. It can be considered as a linear combination of the clear scene and the air-light, as shown in Eq. (1):

$$F^c\left(x\right)=J^c\left(x\right)t\left(x\right)+A^c\left(1-t\left(x\right)\right)$$

where c ∈ {r, g, b}, F^c(x) ∈ R^C is the haze image, J^c(x) is the clear scene radiance, A^c is the brightest point in the sky, and t(x) is the scene transmission with the depth. Since A^c is usually obtained from experience, t(x) has already become the most important research in haze removal. To make the ill-posed problem tractable, the auxiliary depth information is an effective way for solving this problem. However, not all scenes have objective depth information. Different from early algorithms, most existing endeavors are devoted to finding strong assumptions or constraints for overcoming disadvantage influence. So far as known, there are two kinds of statistical methods for estimating the transmission map. The first type is called Contrast prior [1], [2], [3], [4], [5] which has been taken by Tan [2] as the representative method. Contrast [4] and edge information [5] are treated as an effective clue so that optimal transmission map is dynamically calculated. Another class of techniques is Chromaticity prior [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20]. It has been demonstrated to be widely attractive and practical than contrast prior. Assuming that similar color has the same scene depth, color consistency can be regarded as an indication for media transmission map [6]. For instance, Fattal [7] exploits the geometrical model to restore degraded images. This is accomplished via the assumption that the attenuation factor and surface shading are locally independent. It can produce positive dehazing images but complex computation has been introduced. Based on the theory of Color Line, Fattal [8] also uses geometric triangle to express the relationship of local pixels and obtains the scene transmission map. As the most concerned method, He et al. [9] firstly discovers a simplest dark channel prior that a minimum intensity value in a haze free image will be found in at least one color channel. In consideration of this prior, the initial transmission can be easily predicted. Due to the intelligible theory and satisfactory performance, many progressive methods [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23] have been proposed for variety applications on the foundation of DCP. For example, in order to refine the preliminary results, Tarel and Hautiere [10] infers a veil refined by median filter, Jing et al. [11] and He et al. [12] employ bilateral and guided filter to smooth the atmospheric veil's rough estimation, respectively. In view of the boundary constraint theory, Lai et al. [13] and Wang et al. [14] restored better results than traditional dehazing methods. Unfortunately, halo artifacts and color distortion are still introduced in the sky region which does not conform to DCP constraint. As a result, Wang et al. [15] and Luan et al. [16] put forward the methods to predict the dynamic atmospheric light separately. Besides that, ground radiance suppressed-HTM [17], HDCP module [18] and VR module [19] are employed to directly improve DCP prior. Laplacian-based restoration [20] and linear intensity transformation [21] are also adopted to restore the degradation images.

Note that although all of DCP-based methods have obtained better restoration results, these algorithms still have common problems. As a matter of experience, one of the main reasons is that the principle of DCP. It makes the restored image darker and details that cannot be distinguished. More importantly, it is obvious that the process of light scattering not only changes image luminance but also cuts saturation down. Nevertheless, most of algorithms are only based on the luminance characteristics and neglect the effect of the saturation layer of images.

To resolve the aforementioned problems and motivated by HSI color space, we improve the DCP estimation by mixed strategy for dehazing single images. Unlike traditional models for estimating the transmission map according to metabolic brightness, we combine the saturation with luminance layers, and perform to estimate the initial transmission map. Similar to other methods on suppressing the influence of halo artifacts and inaccuracy points in the initial transmission map, a refinement of process is introduced to achieve the adaptive transmission compensation. Considering that the local color priors have an influence on the accuracy of the recovery results, we utilize local color constraint to generate the optimal transmission map based on the calculation of geometric projection. The main progresses of our proposed approach are organized into two parts as follows:

• Different from many others dehazing techniques which only focus on the luminance layer assumption for estimating the initial transmission map, our method emphasize more on the combination of the saturation and luminance layer. As a result, the proposed can produce high quality fog free image effectively which more visually attractive to observers.

• Compared with the color line model often discards patches [8], we assign essential conditions to shrink the patch for finding out the valid points around the center pixel. More importantly, we deploy them as a constraint for edge preserving, which brings down the error rates of Color Line.

Experiments have shown the proposed method performs on par with or better than the advanced dehazing methods. Firstly, we brief dehazing approaches can be treated as a typical situation of image restoration so that our visual features can restore the initial transmission map through the variational framework in Section 2. In Section 3, we presents improved color line model which is dedicated for producing refined and regularized transmission maps for dehazing algorithm in detail. Furthermore, we present plenty of comparison experiments to verify the effectiveness of our algorithm in Section 4. And finally, we conclude the paper and the future work is given in Section 5.