Let PC read your figure

This article is reprinted from the new computer CHIP "No. 5, 2010, the technology and the future" column will in the coming period, progressively introduce new Microsoft technology, we will promptly reprint, to share with you an exciting new technology.

Computers have thought, with human judgement, this seems to be science fiction scenarios.

However, Microsoft Research Asia technical experts are convinced by constant technological innovation, the computer's judgment can achieve even more than human, with regard to the image of to distinguish some of their research results have been beyond our imagination.

Computer operation speed of development by leaps and bounds, and Moore's law continues to play a role in the ongoing, however this is called the computer equipment was more like a "anencephaly", it can only be mechanically repeated some of mankind's instructions in advance, no advanced judgment.

It can only save the photos to us, but cannot help us figure out from the crowd, or DOE, John helps us modify photos but it is difficult to accurately distinguish out of the blue sky and the beach. However this is Microsoft's Visual Computing Group of researchers has quietly changed.

PC can look "which means" health plan

The relationship between the computer and the graphic images are mainly concentrated in two main lines on the development of a computer drawing, which is the use of computers and two-dimensional and three-dimensional drawing software to build a picture, such as a 3-d modeling with light and shadow generated almost exactly the same with real world photos or video; the second is the use of computers to get pictures and video of nature and the use of computers to store, organize, and edit these resources, which is the Visual Computing.

Created

by computer resources because of the source data that will be computer interpretation easier, for example in the Photoshop hand-drawn blue sea and sky, can be on separate layers, select and process all very convenient. But from the nature of Visual resources you get absolutely no use to distinguish between information of source data, so how to make PC can "read" these visual resources, and the significance of recovery, became in each Visual Computing researchers in front of the subject. Read the Visual resources mean similarly divided into multiple levels, great to have the computer automatically distinguish photo car or pedestrian, small-to-picture alignment needs computer can "understand" Visual resources.

In the era before the advent of the Internet, our data has not been so abundant, computer processing of data, more to play the storage, transmission and output roles.

However in the rapid increase of Internet bandwidth, our main information carriers are starting from text into pictures or even video information, along with an image and video capture device for rapid popularization, graphics and video resources to share and Exchange sites on the Internet, the proliferation of Visual resources, explosive growth rely on humans to manually edit the Visual resources add source data information is not realistic, therefore by computer "reading figure" needs become more prominent.

In the traditional areas of digital information analysis, such as military radar, sonar signal or electromagnetic signals, these signals are professional and high-precision equipment, in accordance with uniform standards for the quality of the information collected, is very high, with little interference factors, and these data information content relative to single, the amount of data is relatively small.

Currently, Visual resources on the Internet is exactly the opposite, because of geographical, photography and acquisition equipment, considerabledifferences, these visual resources very disorganized, irregular, and quality are varied, but the scale of these data is the exception. Therefore, the Internet's visual resources analysis to adopt a new point of view, it is necessary to solve the problem of redundant information processing, and to standardize on visual resources, and ultimately to reuniting from these data gets to have valuable information.

Now Visual computing research are two directions of development, the first is based on statistical analysis techniques of Visual resources, this technology mainly rely on the analysis of the characteristics of Visual resources, and apply these characteristics with statistical analysis methods, such as mathematics and extracted eventual application to practical work; another way the rise in the last two years, and the traditional two-dimensional, low dimension means different mathematical analysis, this new analysis of the perspective from the original mathematical sense of the low dimension to a multi-dimensional, using new mathematical perspective in multiple perspectives on looking at the issue again.

With the help of new mathematical models, many seemingly impossible visual computing problems have been solved, such as with this new model, with a mask or sunglasses face can be read and identify your computer. Both techniques have already begun to show out astonishing achievements, we find that your computer is "towards hope illustration students intended" and "picture" of the era.

Album sorting PC do

The popularity of digital camera equipment for our digital photos of the rapid growth in volume, how to sort these photographs became the face of each person.

Latest Visual Computing face classification technology is maturing, and even some manufacturers already applied in the latest photo management software.

The way the human brain with the selected picture

At Microsoft

research in this area as early as 2005 has already begun, Microsoft researchers want the computer to automatically recognize the face of Characteristics, and according to these features automatically complete the Group's operations. Achieving this goal is experiencing the greatest difficulty is how to choose a characteristic points, because the photo is in different scenes and environments to shoot, so even if the same person, his facial features are also because of the light and expression, and so the impact of different factors change, to improve the efficiency of the automatic face recognition, you must find external adverse factors not sensitive characteristics or characteristic composition. Human face identification when using a similar method, the first identification of the human brain can find out a person's key features, while you can ignore all because of external factors and the characteristics of changes occur, this seemingly contradictory mechanisms to help us distinguish neighborhood uncle and aunt Lee Chang, however how can allowing computers also have similar capabilities?

At first, let the computer on which the identification was manually specified, so the specified feature is usually not desirable, would affect the recognition accuracy.

While Microsoft's researchers found that can face with massive database to calculate and optimize these characteristics and characteristic combinations, they will already contain the exact result of massive face database as a sample, using machine learning methods continue to try to construct the facial features and characteristics of the combination and use of this structured data to calculate the recognition accuracy, after such a massive computational statistics, eventually a group you can get the most accurate face recognition feature combination, this set of features on the outside to the disadvantage of factors has the best resistance. Next by considering the actual use of the environment factors, such as the present computer average CPU and memory handling capabilities, and then on the optimization of feature combination to optimize again, finally apply it to the photo finishing software, photo finishing software will use this group to handle all the eigenvalues of the picture, in order to achieve recognition of facial and the automatic grouping of functionality.

By counting the features in the video, you can easily find out on behalf of Lee inside movement of people

With integrated technology to improve the recognition rate

Although grouping of face recognition technology has been a number of manufacturers into practical applications, and Microsoft will soon in the near future in the two products in the use of the technology, but the technology still greatly improve space.

At present based on facial feature extraction combination simply stuck in as far as possible and human approach, far not reach beyond human ability, for example, in recognition of face recognition, computer efficiency will be decreased significantly, and in one side face to more than 45 degrees, the computer would be difficult to associate it with a positive face is classified as a class.

But in contrast, human identification of capability in the face instead of lateral 30 degrees is the best, this shows a human feature extraction or current sea election by a certain feature group.

In addition, compared with the computer, mankind has with associated features the ability to determine, for example, human beings in judge a person, with the current location, each other's hair and clothes will be able to distinguish, even just to see this person's back is able to accurately determine the characteristics of these associations has greatly increased the capacity of human judgement.

In fact Microsoft Research Asia researchers found that when only face without hair, and other related features, human identification skills and computer differences are not large. While the computer is able to extract these associated features, but due to these characteristics, judge the reliability of the information is difficult, so at present with associated feature information for auxiliary judgment of technology still premature.

Currently, Microsoft Research Asia researchers are still working with associated information to refine and improve face recognition capabilities, for example in the picture library to import new pictures of a group of the party, though the picture gallery has a foreign relatives looks and you are very close, but this man appears in the gathering of the probability is very low, with the help of this information can help a computer aided judgment, improving computer to distinguish between the two of you.

The relationship between data like this may and in face of its own characteristics and not related, but it will help to improve the efficiency of face recognition.

In the face of the relevant characteristics of the information, it is not entirely without according to the search, such as with hair that feature data associated to discern, although the hair will constantly change, but combined with probability and statistics, hair still in progress when you face recognition can play a role.

Drawn by PC Digital "impression"

Let the computer read the image, is a great visual computing challenges, in order to solve this problem, use the traditional approach has been very difficult, but nearly two years of adoption from a multidimensional perspective on visual computing problems it has made great progress.

A new mathematical model of change

Face in the picture information effectively crawl out of the most important thing is to

have a proven mathematical models and algorithms will not be very efficient and very accurate to grab the picture information. Currently face recognition of mathematical models and calculations still is mainly based on traditional statistical model, and this model is mainly used for low-dimensional signal processing and interpretation. While the picture or video this high pixel data is a high-dimensional data, with the Visual computing requirements, researchers found that the use of traditional mathematical calculation will have obvious limitations, breakthroughThe difficulty will be very large, in light of this, we must find a new mathematical model.

A few years ago, a Chinese-American mathematician taozhexuan Australia, represented by some mathematicians first to realize that in high-dimensional space, some originally recognized difficult (NP-hard) combination, you can use a series of highly efficient optimization algorithms to solve.

Mayi research group soon realized that these powerful calculation tool is used to resolve the current visual computing challenges, but the final result is very satisfactory. This new model and theory in the last two years led to discussions, and mathematical models and algorithms are constantly being optimized, this new way of thinking is gradually bring Visual technology breakthrough.

Used in high-dimensional space calculation of new ideas, many things have changed. many of the concepts and tools have built in low-dimensional space, while in high-dimensional space, many results with low-dimensional space, on the contrary, low-dimensional space that will be happening in high-dimensional space in general do not occur.

So in the traditional way of low dimensional calculation considers the basic practice of things, even people who do things that can be implemented in high-dimensional space. For example, in the traditional sense, a 70% ~ 80% of the content is highly damaged photos regardless of the computer or who are no longer able to identification, but in this new model, the remaining 20% ~ 30% of the image the amount of data continues to be alarming, it still can be used for precise calculation, therefore, for accurate identification is not a problem.

Visual Computing beyond human

In these mathematical models and calculations based on the idea of Microsoft researchers are continuously optimized and improved technology and start raises some very interesting application.

For example, they are working on trying to be the same person on the network for great photo data import, and through the algorithm on screen elements for analysis to achieve automatic alignment. More amazingly, by optimizing the algorithm, this tool can automatically analyze and identify this person each important organs of features that enable this feature to repair other incomplete pictures, such as no hair photos to remove the hair, eyes front of sunglasses and even be laughing mouth into a smile. This automatic patching technology based entirely on the same individual facial features of data analysis and, therefore, has the extremely high accuracy. Interestingly, the computer can also be a person's facial characteristics for each key commonality extracted and combined generate this "visual impression". Generate virtual impression of photos of data derived from the human mass photo's comprehensive statistics, so even the virtual photos still vivid, and our brain to coincide with this impression.

Microsoft Research Asia also try this technique in the video, the video is actually a continuous playback of picture, consecutively and the associated picture fits new mathematical models of computing needs, through continuous analysis and calculation of the picture, the new visual computing system can identify each picture in the video the similarities and differences, with these data, you can realize many of the original appears to be

unable to complete video editing features, such as the extraction from video in only two movements of people, or to repair old movie film of scratches.

Exciting is that with this emerging technology there is a huge potential value, attracted mathematicians, statisticians and other professionals such as engineers and widespread attention, this technology is rapidly maturing, computational efficiency is improved by leaps and bounds, is estimated to require 3-5 years or so, this technique can be to the public.

Microsoft's characteristic Visual Computing

Similar image search

In addition to the face, all pictures are there features, Microsoft Research Asia, researchers are trying to search for keywords and picture characteristics to achieve a more precise search.

For example, the application must be (Bing) image search in Visual Computing technologies, we first of all, you can use keywords to search for "Apple", find all the pictures associated with Apple, we also can click on the Red Apple pictures next to find similar pictures link to find all the pictures of the Red Apple.

First cases of the use of "CHIP" in the keyword search, and then select similar picture features that Visual computing technology can be similar pictures to find out

Microsoft Research Asia of the researchers as a single object, outdoor scenes, one major types to differentiate between the picture and the picture setting for each category a group characteristic value, through extraction and on the characteristic value that helps users finally found similar pictures.

365 Sky Dream

One year of the number of blue sky symbolizes our living city air quality standards in the digital world, Microsoft Research Asia, they have driven away the mist, through Visual Computing to fog technology, any pictures or vide

o of mist can be removed.

The difficulty lies, how to let the computer know that photos or video of interference in the fog, the computer needs to be able to distinguish the mist and a picture of a white background.

Microsoft Research Asia, researchers found respectively red, green, blue, 3-color black channel, in red, green, blue, 3 channels in each pixel, peripheral range is set to a temporary domain, in the temporary domain most black-point instead of the original pixel point, this kind of "pro domain take small" process can get black channel. The resulting interesting phenomenon is that if you are a no fog of colorfulPictures, black channel is black, not see what content, while the fog of pictures of the black channel is not so black. This makes it possible to distinguish whether the picture with fog. Through the following series of formula calculations, Microsoft Research Asia engineers not only can you get rid of the fog, the fog can also be extracted separately, because the distance fog fog light, near, we can also get this photo of depth information.

By placing pictures R, G, b 3 channels into black channel, you can identify whether a fog in photograph exists

At present the defog Visual algorithm efficiency is very high, with a higher PC, with the help of this algorithm can achieve when shooting video in real-time to fog.

High speed smart select

Do not know if you would remember, CHIP was introduced then Office2010 picture editing tools has added the function of automatic selection, this feature can help us automatically select pictures of prospects and get rid of the background.

In fact in this technology in Microsoft Research Asia has become more powerful.

We no longer need a little choice of foreground and background, you only need to use a similar Brush tool in the picture quickly flashed to qu

ickly build image selection. More importantly, this selection algorithm after special optimization, even in a dozens of megabytes or even hundreds of megabytes of photo on it, or you can instantaneously, far faster than the current mainstream graphics editing software of similar functionality, but also on the edge xinidu, Microsoft's Select tool performance is also very good.

New image selection technology easy to operate, extremely high accuracy, and speed

——————————————————————————————­————————————

Welcome micro- concern soft Asian Institute micro-Bo: http://t.sina.com.cn/msra