Single Image Dehazing¶
- Single Image Dehazing
Introduction¶
Image captured in bad weather often suffers the loss of visibility due to the existence of fog or haze in the atmosphere. This is because the light reflected from target objects will get attenuated through the haze. What we get from our camera is a mixture includes original colors of objects and degradation effects of haze. Which is, obviously, inconvenient for our subsequent steps of image processing and computer vision.
In this page, we summarize some popular algorithms on single image haze removal, and classifies them into several categories. In the last, we propose our hardware design.
Basic Knowledge of Image Dehazing¶
In computer vision and image procession, the physical model of a hazy image is:
In this function, I stands for the image to be processed, while J means the original scene without being hazed. As is known to all, the original radiance of objects suffer attenuation when it goes through fog or haze, thus we multiply J with transmission map t, which means the proportion of radiance reaching the camera without the effect of haze. From the definition, we can simply conclude that the value of t lies in [0,1]. Another important part of a hazy image is composed by atmospheric light A, it compensates for the loss of original radiance.
The goal of image haze removal is clear now. From a single hazy image, we need to evaluate the value of transmission map t and atmospheric light A. Then we can restore the scene by equation:
Current Research¶
Since we only have one hazy image and one physical model, it is difficult to estimate proper transmission map and atmospheric light. However, much academic work has been done to compensate for the scarce information that a single image can provide. Some methods proposes new prior of hazy image, among which the dark channel prior is widely used today. Some methods considers various characteristics of a image, such as image contrast, image frequency, image color distortion. etc, and then search a balance among these factors to produce a vivid dehazed image. Others provide different physical models, such Markov field models, using theory of probability to realize image haze removal.
The most widely used prior is dark channel prior:
Dark Channel Prior¶
He's Dark Channel Prior¶
★★★★★
Dark channel prior is widely used by researches after it has been proposed. Since the original prior needs postal step--soft matting, which is very time-consuming, researchers have proposed many other methods combining the dark channel prior and novel subsequent processing.
Combines Dark Channel Prior with Guided Image Filter¶
★★★★★
Guided image filtering is proposed by the same author of dark channel prior.
Combines Dark Channel Prior with Associate Image Filter¶
★★★
Associate filter can transfer the structures of a reference image and the grey levels of a coarse image to the filtering output.
Other Priors of Hazy Image¶
Fattal¶
★★★★
Resulting shading and transmission functions are locally statistically uncorrelated.
Tan¶
★★★★
Based on two observations:
1. Restored image have a larger contrast then hazy image.
2. Airlight whose variation mainly depends on the distance of objects to the viewer, tends to be smooth
Han¶
★★★★
Han proposed a two-peak channel prior based on He's dark channel prior to improve the robustness of validity.
Optimization Methods¶
Different from methods listed above, optimization methods seeks to improve various qualities of a hazy image, such as image contrast, image context, image color information, etc. It's obvious that we cannot improve all the qualities to the best simultaneously, however, we can search for a optimal transmission point. Generally, these methods propose a cost function, which considers 2 or 3 quantitative characteristics of a image, then search for the point that minimize this cost function.
Boundary Constraint and Contextual Regularization¶
★★★★★
This paper considers the boundary of transmission map, then try to save as much contextual information as possible. All of these are achieved by minimizing a cost function considering these two factors.
Image Contrast and Color Information¶
★★★★★
This paper builds a cost function of the term image contrast and information loss. By minimizing this cost function, the contrast of image get enlarged, while most of the color information get stored.
Other Methods¶
Tarel¶
★★★★
Estimating atmospheric veil, then using corner preserving smoothing.
Fast Image Dehazing Using Guided Joint Bilateral Filter¶
★★★★★
Derived from Tarel's method, then use guided joint bilateral filter to track the edge information of images.
Kratz's Probabilistic Method¶
★★★
This paper introduce a novel probabilistic method that fully leverages natural statistics of both the albedo and depth of the scene to resolve image ambiguity. The key idea is to model the image with a factorial Markov random field in which the scene albedo and depth are two statistically independent latent layers.
Hardware Implementation of Image Haze Removal¶
Less papers are found in this area compared to software algorithms .
Hardware Implementation with Dark Channel Prior¶
★★★★★
This paper introduces a hardware method of image haze removal, based on dark channel prior.
Our Proposed hardware method¶
Based on an optimization method, we have proposed a hardware design for image haze removal.
Other Recourses¶
Some images needs to be dehazed can be downloaded in this page. They can be used in Matlab and Verilog experiment. Also, You can see some brief presentations of several articles made by myself.
D-Hazy Dataset¶
D-Hazy Dataset is a dataset that contains 1400+ pairs of images with ground truth reference images and hazy images of the same scene, which is convenient for Quantitative Evaluation.
