(han) Single image haze removal using content-adaptive dark channel and post enhancement

use content to adapt to dark channels and post-enhanced single image fog

absrtact : As a challenging problem, image fog plays an important role in the application of computer vision. The dark channel Priori method has been extensively researched for image de-fog because of its simple and efficient characteristics, however, there are some problems with this method: Supersaturation, artifacts and darker appearance. In order to solve these problems, this paper presents a method of single image de-fog, using the method of dark channel and post-enhancement of content adaptation. The main contributions of this study are as follows: First, an associated filter is proposed for efficient dark channel calculation, which can convert the structure of the reference image and the gray level of the coarse image to the output of the filter; second, the dark channel confidence based on the content of the image is used to limit the dark channel. Finally, a post-enhancement method is designed to map the brightness of the fog-free image and preserve the local contrast. The experimental results show that the proposed method can significantly improve the visibility of fog images. 1. Introduction

Due to atmospheric absorption and scattering, the irradiance received by the camera from the scene point is attenuated along the line of sight, and the incident light is mixed with the air light. This phenomenon, known as haze or fog, can significantly reduce the visibility of the scene. Most computer vision applications, such as image segmentation and object tracking, are often affected by poor visibility of fog images [1-3]. Therefore, mist removal is very much needed in many practical applications. Typically, the correlation between fog and scene depth is high. [4-8] because it is difficult to estimate depth of field from a single image, an earlier method of de-fogging typically requires multiple input images or additional information. However, in many cases it is not possible to fully meet these requirements, so people based on strong priori or assumptions, a single image method is proposed [9-13].

Recently, the single image of the fog has progressed significantly. Some methods increase the visibility of the image by contrast enhancement [11,14-17]. These methods generally do not take into account the depth of the scene, so they are essentially not able to remove haze as the depth of scenes changes. For example, the Tan method [11] significantly improves visibility by maximizing local contrast, but the results often look unnatural due to over-enhancement. Some other ways to remove fog are through image recovery. The Fattal method assumes that the transmittance and surface shading are locally unrelated, although the method is physically sound, but does not handle severe fog well [10]. He Keming and others [9] proposed dark channel priori, that is, the outdoor non-fog image of most local tiles contain some pixels, these pixels in at least one color channel strength is very low, so they simply reduce the dark channel to zero mist.

The Dark channel priori is simple, but it is effective for single image fog. Based on the work of He Keming and others [9], many methods of using dark channels have been proposed [9,13,14,18]. However, it is possible that some areas of the image do not meet the dark channel preference, and the calculation of dark channels is not an easy task. As a result, these methods may suffer one or more of the following issues. First, some methods often lead to color supersaturation because they do not meet the specific restriction requirements: the resulting dark channel should not be brighter than the smallest color channel. Second, these methods often introduce artifacts into specific smoothing regions in which dark channel priors are unreliable. Finally, after the atomization is removed, the restored fog-free image usually looks dark because the irradiance from the field is attenuated during its propagation.

To solve the above three problems, we propose a method of using content to adapt to dark channel and post-enhanced single image blur removal. The main contribution of this work is as follows: first, in order to obtain the dark channel effectively and efficiently, we propose the correlation filter, which can transform the structure and gray level of two input images into the filter output respectively. Secondly, we propose dark channel confidence to limit the dark channel based on the content of the image. Finally, inspired by the features of the recovered fog-free image, we designed a post-enhancement method to map brightness and preserve local contrast. Compared with the prior art, the experimental results show that the proposed method produces satisfactory results on different fog images.

The remainder of this article is organized as follows. The following section describes an observation of problems with existing methods. The 3rd section gives the definition of correlation filter, the concept of dark channel confidence, and the process of using content adaptive dark Channel method to fog. The 4th section discusses the characteristics of the recovered fog-free image, describes the post-enhancement method, and the 5th section deals with some existing technical methods for comparative experiments. Finally, the 6th section concludes. 2. Observe

The atmospheric scattering model, usually simplified to (1), is widely used to remove fog [4,19,20]. Using the Dark Channel transcendental method, the transmittance t (x, y) t (x, y) and the recovered fog-free images are computed by (2) and (3) respectively. The restored fog-free images are prone to supersaturation, artifacts, and low brightness overall, as shown in Figure 1. In this section, we present a look at three questions.
Ic (x, y) =jc (x, y) tc (x, y) +ac (1−TC (x, y)) (1) i^c (x, y) = J^c (x, y) t^c (x, y) + a^c (1-t^c (x, y) \tag{1}
TC (x, y) =1−minc∈{r,g,b} (min (i,j) ∈ω{r,g,b} (Ic (x, y) ac) =1−idark (x, y) ac (2) T ^c (x, y) =1-\min_{c \in \{r,g,b\}} (\min_{(i , j) \in \omega\{r,g,b\}} (\frac{i^c (x, y)}{a^c})) =1-\frac{i_{dark} (x, y)}{a^c}\tag{2}

Jc (x, y) =ac+ic (×,