[Original: https://www.cnblogs.com/molakejin/p/5708883.html ]
1: The best paper from Jane to Beauty (Author: He Yuming Visual Computing Group)
[Visual robot: I feel that it is important to learn his classic algorithm, but his problem-solving ideas are also very worthwhile for us to learn]
It was the morning of April 24, 2009, and I received an unusual email. The senders are the chairmen of CVPR 2009, and they said that my article won the Best Paper Award of CVPR 2009 . I read this email over and over to make sure I didn't misunderstand. It's an incredible thing.
Dehazing results of Beijing smog photos
The Chinese name of CVPR is the Conference on Computer Vision and Pattern Recognition, which is one of the top international conferences in the field of computer vision. This year's CVPR received about 1,450 submissions, of which 393 were accepted, an acceptance rate of 26%. Only one article was selected as the best paper of the year. This is the first time a Chinese has won this award since CVPR was founded 25 years ago. This article was completed when I was an intern in the Image Computing Group of Microsoft Research Asia, and it was also the first paper I actually wrote.
Simple and effective image dehazing technique
The problem studied in this paper is image dehazing technology, which can restore the color and visibility of the image, and can also use the concentration of fog to estimate the distance of objects, which have important applications in computer vision (such as three-dimensional reconstruction, object identification). But people have not found a simple and effective way to achieve this goal before. In this paper, we find a very simple, even surprising statistical law, and propose an efficient dehazing method.
Unlike previous methods, we focus on the statistical features of haze-free images. We found that in a haze-free image, every local area is likely to have shadows, or something that is pure color, or something that is black. Therefore, every local area is likely to have at least one color channel with very low values. We call this statistical law Dark Channel Prior. Intuitively, Dark Channel Prior believes that there is always something very dark in every local area. This law is very simple, but it is the basic law of essence in the dehazing problem we study.
Since fog is always off-white, once an image is affected by fog, things that should be dark can turn off-white. Not only that, according to the formation formula of physical fog, we can also judge the concentration of fog according to the grayness of these things. Therefore, our proposed Dark Channel Prior can effectively remove the influence of fog, and at the same time use the concentration of the object to estimate the distance of the object.
Inspiration from computer games
The idea arose from two serendipitous observations.
The first observation comes from a 3D game. This game has many scenes with fog, but these scenes are fictional and unreal. Computer-generated 3D images can be quite different from the statistical regularities of natural images, but the human visual system can still sense the fog that exists in virtual images. This leads me to believe that the human visual system must have an efficient mechanism for perceiving hazy images, and that this mechanism must be different from existing dehazing methods. The dehazing methods proposed by the predecessors all focus 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 fog in virtual scenes, indicating that the human eye must be using something other than contrast to perceive fog. So I think there must be something closer to the essence in this question that people haven't discovered.
The second observation comes from the study of previous dehazing methods. The previous most effective dehazing method was proposed by Fattal in the Siggraph article "Single Image Dehazing" in 2008, which is our primary goal of surpassing. In the comparisons presented in this post, I found that a method called Dark Object Subtraction sometimes works better. This method utilizes the darkest point in the whole image to remove globally uniform fog. This method is more efficient if the fog is indeed uniform. The downside is that it can't handle uneven fog, which is the hard part in the dehazing problem. So the natural idea is to locally use Dark Object Subtraction to process images. And it just so happens that it doesn't need to use contrast, which shows that it is fundamentally different from the previous method. Surprisingly, in numerous experiments, I found that such a simple idea worked very well.
But the most important ideas in our paper were formed after I started writing the article. In the first few drafts of the article, Sun Jian, my mentor in the image computing group, kept asking me what was the essential reason for the success of our method, and what was behind the "insights" that we did not fully understand. Although we have a very simple method and beautiful experimental results, we are not able to convince people of the effectiveness of this method. This is because we have not yet come up with a rationale. With this question, I returned to experimentation and observation. I found that since a large number of experimental results have confirmed that the partial implementation of Dark Object Subtraction is successful, it means that there are indeed dark objects in each part of the image after dehazing. That is to say, behind the success of this method, there is actually a statistical law about fog-free images. My mentor Sun Jian asked me to research a database of fog-free images first. Through a large number of experiments, we found that this statistical law exists objectively. This is what we propose as Dark Channel Prior.
This is the first paper I wrote
In 2007, I graduated from the basic science class of Tsinghua University 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 level, I did not systematically study related knowledge of computers. Out of interest, I took some related courses in computer graphics and graphics. But in the early days of my internship at Microsoft Research Asia, these basic courses were far from adequate for the research work I faced. The lack of background knowledge made it difficult for me to get started. When reading articles, I often do not know which methods are used by everyone and which are the authors' contributions. For me, everything I see is new.
During the interview, my mentor Tang Xiaoou told me that he didn't care that I didn't have relevant background knowledge, because all relevant things can be learned. In the first year of internship at Microsoft Research Asia, I did several different projects under the guidance of mentor Sun Jian. Although none of them were successful, I learned a lot from them. Among them, the image matting problem (extraction of semi-transparent object boundary), which I spent a lot of time researching, was very helpful to this article. When I first started to study dehazing, I found that the equation of fog is very similar to the equation of matting, and the matting framework I studied before can help dehazing. Using this framework, I just had to find a way to estimate the fog concentration locally. This framework allowed me to concentrate on finding such a method and finally came up with Dark Channel Prior.
Dehazing results of smog photos in New York and Beijing
Even with ideas and experimental results, writing an article for the first time made me feel very difficult. I often get caught up in my character fighting with myself. After writing each paragraph, I often ask myself if this is the case, and if there are any loopholes in it. I also ask myself, if I am a judge, or a reader, can I understand this article, and how can I write to make my thinking flow more smoothly. In the midst of such struggles, a draft often takes days to write. Even so, the first few drafts were far from satisfying Sun Jian. At the beginning, he only gave me suggestions on the structure of the article, ideas and viewpoints, but did not specifically revise my article. So I went back and continued arguing with myself. But whenever I convince myself, Sun Jian can always raise new questions. In this cycle, one day Sun Jian finally said that the article had been written well, and he began to revise it in detail. It is precisely such demanding requirements that there will be subsequent high-quality articles.
The journey of the avenue lies in simplicity
All three reviewers of our article gave 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 it to be true. Another judge felt that few articles have achieved such a large improvement in experimental results in such a simple way. One of the judges even implemented our method himself and confirmed that it works. Sun Jian said that it was a joy to read such a review result. Mr. Tang believes that the success of this article lies in three aspects. First, the method is very simple; second, it gives good results for a very difficult problem; third, a basic natural law has been discovered and applied to practical problems. The audience also gave it high marks after the presentation in Miami. They told me it was one of the funniest talks at CVPR this time.
A researcher at the meeting said that the best ideas are often those that seem simple, but everyone will think that no one has thought of them. And our idea just fits this point. The first sentence of our paper abstract says, "We propose a simple and efficient method". Perhaps, this is the best summary of our work - simple, is beautiful.
about the author
He Yuming: Intern in the Visual Computing Group of Microsoft Research Asia, currently studying at the Information Engineering Multimedia Laboratory of the Chinese University of Hong Kong, and graduated from the Basic Science class of Tsinghua University. He is one of the winners of the Microsoft Junior Scholars Scholarship in 2006 and the No. 1 Scholar in the Guangdong Provincial College Entrance Examination in 2003.
Two: Fast implementation of advanced image dehazing algorithm
http://blog.csdn.net/huixingshao/article/details/42834939
This is a blog post that introduces He Yuming's dehazing algorithm in detail, and gives the implementation program. It seems that the source code does not
Three: Single Image Haze Removal - CVPR'09 Best Paper
http://blog.csdn.net/abcjennifer/article/details/6662706
Matlab and Opencv code given by CSDN famous blog Rachel-Zhang and detailed introduction
Four: He Yuming Blog
http://research.microsoft.com/en-us/um/people/kahe/
Classic: CVPR best paper award in 2009 Classic Dehazing Algorithm Single Image Haze Removal Using Dark Channel Prior. It is the first time that a Chinese person has won this award since CVPR was founded 25 years ago.
His paper on dehazing and improvements continued until 2013 at: http://research.microsoft.com/en-us/um/people/kahe/cvpr09/index.html
方向：deep learning for visual recognition, including image classification, object detection, and semantic segmentation.；