How do I find another picture through a picture?

Absrtact: If the user gives a picture, Baidu map will judge whether there is a face, if there is, Baidu map in a similar image search, while the whole network looking for similar portrait. How do I find another picture through a picture? For search

If the user gives a picture, Baidu map will judge whether there is a face, if there is, Baidu map in a similar image search, while the whole network looking for similar portrait.

How do I find another picture through a picture?

For search engines, there is no essential difference between finding the inner link between the images and the Common keyword search--all through the matching of key features, according to certain logical rules. However, the difference is also obvious, with a picture of the input launched by the search, there are many possible search implications.

For example, a picture may include both landscapes and people, and what users want to look for is a landscape, a layout-like picture, or a similar person? Google to map the function, can even guess according to a picture of the shooting, but does not try to the image of the exact match. Most search engines do.

And most of the search engine as input, such as TinEye (2008 on line), Sogou map (2011 on-line), and so on, is essentially a picture of the approximate copy detection, that is, the search looks almost exactly the same picture. The introduction of the 2010 Baidu Map (shitu.baidu.com) is also the case.

After more than two years of silence, Baidu Map began to explore the other direction.

Last week's Baidu annual meeting, Li specifically mentioned Baidu map: "To map the accuracy of the map from 20% to 80%." However, compared to the previous, the ability of Baidu to find similar images does not appear to be significantly improved, so where does the change come from? Robin Li attributes this apparent elevation to a newly-launched face recognition search.

And before the difference is, if the user gives a picture, Baidu map will judge whether there is a face, if there is, Baidu map in the similar image search, while the whole network to look for the appearance of similar portrait.

The new added technology in short, first is the face detection and extraction of feature expression, and then the database comparison, and finally according to the similarity of the order returned results. In fact, face detection is not a new technology, the relevant research has 30 years of history, but until the end of last year, Baidu decided to promote the implementation of the technology.

Naturally there is a strategic dimension to consider. Two years ago, Robin Li's judgment of the future of the reading age; At last year's KDD conference, Li's nine challenges to solve technology, content-based image search technology is listed in the third; realistic level of Baidu's picture-related products, cloud albums and so on Have a demand for this technology.

Baidu also hopes to use this approach, mining the links between the pictures, and thus stimulate two of times browsing. However, to turn the idea into reality, Baidu must solve at least two problems: first, the algorithm, and the second data.

--Algorithms. The same is based on the image of the search, face recognition and map search is not the same. Baidu Senior Engineer Taugi told the story, Baidu face recognition first does not pay attention to the complete image structure, followed by color also has no meaning. The most important feature expression comes from facial texture and some reprocessing.

The specific algorithm, as a trade secret, is difficult to disclose to the outside world. However, the data show that LBP is a popular feature extraction method in the face recognition algorithm, that is, by comparing the gray value of 8 adjacent pixels around the pixel and the center gray value, a eight-bit code is obtained and then classified according to the coded histogram.

The factors that affect the effect of the algorithm may include preprocessing, feature selection, feature point positioning accuracy, classifier design and post processing, as well as various methods of fusion, threshold selection and so on.

In a picture, the face has at least 40x40 pixels (about a fingernails size) to be treated as a valid recognition object. If a picture of more than one portrait, the current Baidu's solution is only to identify the largest size of that, the future Baidu will provide focus selection function, users can determine the search target.

--Data. If the improvement of the algorithm is sufficient condition, the data processing is the necessary condition. From the perspective of machine search, light, posture, expression, angle and other factors, are influential factors, the so-called "laugh and not laugh, are different." So the larger the amount of data used to train, the better the inclusiveness of change.

When the face recognition search is actually used, the more data the same face accumulates, the better the support can be in matching the contrast. To this end, Baidu needs to be the entire network of tens of billions of images extracted, and then no face of the data removed, and then massive scan again, to establish the index as efficient as possible.

The introduction of face recognition into the search engine is bound to aggravate the external concerns about privacy leaks. Baidu stressed that face search will only be in the open information range, closed personal albums will not be touched.

This also involves the concept of accuracy rate, recall rate and so on. Suppose the database has 100 of Andy Lau's picture, with Andy Lau's head launched a search, the first 50 results in 40 is the real Andy Lau, then the first 50 recall rate is 40%, the accuracy is 80%. Similarly, if the database only two pictures of passers-by a, when using the head of a passer-by to launch a search, the first 50 may only 1 real passers-by a, then the first 50 recall rate is 50%, the accuracy is 2%.

As celebrities in the online photos more, can be perceived by the user of the accuracy rate is much higher than the average, and the recall rate is not perceived by the user. According to reports, face recognition search technology will be in the Baidu Cloud album further integration, to help users to establish links between pictures, and the future does not exclude the possibility of open APIs.

In addition to the above technical discussion, from the research and development system of Baidu, face recognition search is also quite representative.

This has invested dozens of engineers, Baidu Basic Technology Department of the Multimedia department responsible for the core algorithm project, last November late December, the end of the year has been online operation. This speed is not common in Baidu, so some people are half joking that this is Baidu recently "change Style" campaign results.

It is noteworthy that Baidu's newly established basic technology department. The department is led by Baidu Chief scientist Wang and reports directly to Li. Some Li Yanhong are said to be personally involved in a number of key projects. Natural language processing, Internet Data Mining, multimedia, recommendations and personalized technology research and development, are responsible for the Baidu Basic Technology department.

Or in the Baidu Annual meeting, talk about the recent challenges, Li said Baidu will not ignore the value of the channel, but technology is the key to the future industry. "Believe in the power of technology, the future is in our hands," Robin Li issued such a call. Obviously, Baidu's "variation" is inevitable to be staged.