He Cai Classic de-Fog algorithm

He Cai Classic de-Fog algorithmOne: Best paper from Jane to the United States (He Cai Ming Vichier)

"Visual robot: Personal feeling learning his classic algorithm is very important, but his problem-solving approach is also very worthy of our study."

It was the morning of April 24, 2009, and I received an unusual e-mail. The sender was the chairman of CVPR 2009, who said that my article won the CVPR 2009 Best Paper Award (Paper Award). I read the e-mail repeatedly to make sure I didn't understand the error. This is really an incredible thing.

Fog results from Beijing haze Photos

CVPR's Chinese name is a conference on computer vision and pattern recognition, one of the top international conferences in the field of computer vision. This year's CVPR received about 1450 submissions, of which 393 articles were received with a reception rate of 26%. Only one article was selected as the best paper of the year. This is the first time that CVPR has won this award in 25 years since its founding in China. This article was completed during my internship at Microsoft Research Asia Image Computing group and was the first paper I wrote in my personal true sense.

Simple and effective image de-fog technology

The problem of this paper is the image de-fog technology, which can restore the color and visibility of the image, but also can use the concentration of fog to estimate the distance of objects, these are important applications in computer vision (such as three-dimensional reconstruction, object recognition). But people have yet to find simple and effective ways to do this. In this paper, we found a very simple, even surprising statistic, and proposed an effective method of de-fogging.

Unlike previous methods, we focus our attention on the statistical characteristics of fog-free images. We found that in a fog-free image, every local area is likely to have a shadow, a pure color, or something black. Therefore, every local area is likely to have at least one color channel with a very low value. We call this statistic law dark Channel Prior. Intuitively, Dark Channel Prior thinks that there is always something dark in every local area. This rule is very simple, but in our study of the problem of fog is the essence of the Basic Law.

Since the fog is always white, once the image is affected by fog, the things that should be very dark will become gray. Not only that, according to the physical fog formation formula, we can also be based on the gray degree of these things to determine the concentration of fog. Therefore, we propose that the dark Channel prior can effectively remove the effect of fog, and also use the concentration of the object to estimate the distance.

Computer games Bring inspiration

The idea came from two random observations.

The first observation comes from a 3D game. This game has a lot of scenes with fog, but these scenes are fictitious things. Computer-generated 3D images will be very different from the statistical laws of natural images, but the human visual system can still feel the fog in the virtual image. This makes me believe that the human visual system must have an effective mechanism to perceive foggy images, and this mechanism must be different from the existing method of de-fogging. The previous approach to fog has focused on the contrast of the image, but the statistical law of the contrast between the virtual scene and the real scene will be very different. The human visual system is still able to perceive the fog in the virtual scene, stating that in addition to contrast, the human eye must still use something else to perceive the fog. So I think there must be something closer to nature that people have never found in this problem.

The second observation comes from the study of the methods of the previous fog. The most effective way to de-fog before is fattal in the 2008 SIGGRAPH article "single Image dehazing", this article is our first goal to surpass. In the comparative results given in this article, I find that a method called dark Object subtraction sometimes has a better effect. This method uses the darkest point of the whole map to remove the global uniform fog. If the fog is indeed homogeneous, this method will be more effective. The disadvantage is that it cannot handle uneven fog, which is the difficulty in the fog problem. So the natural idea is to manipulate the image locally using dark Object subtraction. And it happens that this does not need to use contrast, indicating that it has an essential difference from the previous method. Surprisingly, in a lot of experiments, I found that the simple idea, the effect is very good.  

But the most important ideas in our paper are formed after I have written the article. In the first few drafts of the article, I have been asking the mentor Sun Jian of the image Computing group what is the essence of our approach to success, and what we do not have a thorough understanding of "insight" behind it. Although we have very simple methods and beautiful experimental results, we cannot convince people of the effectiveness of this approach. This is because we have no reason to speak of. With this problem, I went back to the experiment and observation. I have found that since a large number of experimental results confirm that the practice of locally doing dark object subtraction is successful, it means that every part of the image after the fog is actually a dark object. In other words, behind the success of this method, there is actually a statistical law about the fog-free image. My mentor Sun Jian let me go first to study a database of fog-free images. Through a large number of experiments, we found that the statistical law is objective. This is the dark Channel Prior we are proposing.

This is the first paper I wrote.

In 2007, I graduated from Tsinghua University with a basic science class and then studied at the Chinese University of Hong Kong. The major courses in the basic science class are mathematics and physics, so at the undergraduate stage, I did not systematically study computer-related knowledge. Out of interest, I took some courses on computer graphics and images. But in the early days of entering Microsoft Research Asia Internship, these basic courses are far from enough to cope with the research work I am facing. The lack of background knowledge has made it difficult for me to get on the road to getting started. In reading the article, I often do not know which is the method that everyone uses, which is the author's contribution. For me, everything I see is new.

During the interview, my mentor, Tang, told me that he didn't care that I didn't have the relevant background knowledge, because all the relevant things can be learned. In the first year of entering Microsoft Research Asia Internship, I have done several different subjects under the guidance of mentor Sun Jian, although none of them have succeeded, but learned a lot of knowledge from them. Which I spent a lot of time studying the image matting problem (translucent object boundary extraction), this article has a great help. When I first studied the fog, I found that the fog equation was very similar to the matting equation, and the matting framework I studied earlier could help with the fog. Using this framework, I just need to find a way to estimate the concentration of fog locally. This framework allows me to concentrate on finding such a method and finally proposes the dark Channel Prior.

Haze Photos of New York and Beijing Haze results

Even with ideas and experimental results, the first time I wrote an article made me feel very difficult. I often fall into the role of myself and my own quarrel. After each paragraph is written well, I often ask myself whether this is the case, there is no loophole. I will also ask myself, if I am a judge, or a reader, then I can read this article, how can I write to make the idea more fluent. In this struggle, a draft is usually written for several days. Even so, the first few drafts were far from satisfying Sun Jian. At first, he only gave me advice on the structure, ideas and opinions of the article, instead of specifically revising my article. So I went back to arguing with myself. But whenever I convince myself, Sun Jian still always raises new questions. In such a cycle, and finally one day Sun Jian said the article has been written well, he began to modify the specific. It is this stringent requirement that there will be later high-quality articles.

the trip to the boulevard lies in Jane.

The top three reviewers in this post have given the highest ratings. They think our method is simple and effective. One of the judges said the idea of Dark Channel prior sounded incredible, but we proved its authenticity. Another judge thinks few an article can make such a big improvement in the results of the experiment in such a simple way. One of the judges even personally fulfilled our approach and confirmed that it was possible. Sun Jian says reading the results of such a review is a happy thing to do. And Tang said that the success of this article lies in three aspects. First, the method is very simple; second, for a very difficult problem, give a very good result; Thirdly, we find a basic law of nature and apply it in practical problems. At the end of the speech in Miami, the audience also gave a very high rating. They told me that this was the most interesting speech on the CVPR.

One researcher at the conference said that the best idea was often those that seemed to be simple, but said that everyone would think how nobody thought about it. And our idea coincides with that. The first sentence of our abstract is to say, "We have proposed a simple and effective method." Perhaps this is the best generalization of our work-simple, beautiful.

Author Introduction

He Cai: Microsoft Research Asia Vichier Intern, is currently studying at the Chinese University of Hong Kong Information Engineering Multimedia Laboratory, graduated from Tsinghua University in basic science class. He was one of the 2006 Microsoft Scholar Scholarship recipients, but also the 2003 Guangdong Provincial College Entrance examination champion.

Second: Fast implementation of advanced image de-fog algorithm

http://blog.csdn.net/huixingshao/article/details/42834939

This is a detailed introduction of He Cai fog Algorithm blog post, and give the implementation of the program, it seems that the source code is not

Three: Single image Haze removal (image de-fog)-cvpr ' best Paper

http://blog.csdn.net/abcjennifer/article/details/6662706

CSDN Rachel-zhang the MATLAB and OPENCV code and detailed introduction

Four: He Cai-Ming Blog

http://research.microsoft.com/en-us/um/people/kahe/

Classic: CVPR Best Paper Award in 2009 Classic de-Fog algorithm single Image Haze removal Using Dark Channel Prior. For the first time in 25 years, CVPR was awarded this award by the Chinese.

His papers on the fog and the improvements that lasted until 2013 at: http://research.microsoft.com/en-us/um/people/kahe/cvpr09/index.html

Direction: Deep learning for visual recognition, including image classification, object detection, and semantic segmentation.;

Resources: Paper, code

Updated: 2015

He Cai Classic de-Fog algorithm