In recent years, breakthroughs have been made in computer graphics and image processing technology, and more and more amazing graphics processing software has emerged in personal PCs. With the latest research in the field of mathematics, personal computers have begun to learn to "see" Picture, read text, identify buildings.
The traditional PC image recognition technology is mainly based on statistical principles, which mainly rely on analyzing the characteristics of visual data, and extracting these characteristics with the help of statistical modeling and other mathematical analysis methods, and finally applied to the actual image processing. This image recognition technology is still the current mainstream, widely used in OCR text recognition, face recognition, image processing and other fields. However, this traditional mathematical analysis method has many limitations, such as high requirements for the quality of pictures. This problem cannot be improved until the emergence of new mathematical models. In the "Fashion Technology" column of CHIP in May 2010, we have introduced to you a new development of a PC image recognition technology. In 2009, some mathematicians represented by Chinese Australian mathematician Tao Zhexuan first discovered some NP-hard combinatorial problems that were originally recognized as difficult in high-dimensional space, which can be solved by a series of efficient convex optimization algorithms. The resulting mathematical model can be used to solve the problems faced by current visual computing, and the final calculation results are very ideal.
Researchers at Microsoft Research Asia used this mathematical idea to make a big breakthrough in the field of image recognition. Faces with masks or sunglasses using this new mathematical model can even be read and recognized by a PC. Recently, researchers at Microsoft Research have made new progress in this technical field again. They have allowed PCs to read "understand" buildings, or have the ability to "recognize" words, and correct distorted or deformed words.
Reading pictures starts with understanding the structure
The traditional two-dimensional image recognition technology relies more on image feature points to work. It first obtains the most representative points in the image through statistical methods, and then tries to find these features when encountering new images. Point, and compare the found points with the original feature points. This technique often works well when the picture quality is excellent and there is no distortion. But the reality is that when we take pictures, due to many reasons such as light, location, etc., we can't finally obtain qualified images, which greatly limits the development of this image recognition technology.
Researchers at Microsoft Research Asia tried to use high-dimensional mathematical models and optimization tools to solve this problem. To understand it simply, the high-dimensional mathematical model adopts the matrix model, which can help us look at the objects in the image in a holistic concept, instead of only obtaining local feature points like traditional techniques, which is more like seeking the whole of the objects in the image Symmetry and regularity. For example, the windows of a typical building are straight and rectangular, and the table is always boxy with 4 legs. With these rules, even if the picture can only provide limited information, the PC can more easily identify the objects in the picture. In the high-dimensional mathematical model, the data input at each point can be used to predict a certain regularity, so this high-dimensional image recognition technology can use almost every pixel in the picture to obtain the overall rule of the object in the image Structure, which means that often only a small part of the picture can complete the correction and recognition of the objects in the image. For example, in traditional image recognition technology, the 100 & TImes; 100 image area often does not provide much feature point data, while in high-dimensional image recognition technology, this means that nearly 10,000 pixels can be used to obtain the image. Rule structure information.
Read the picture from a human perspective
Recognizing the surrounding environment and scenery with the help of regularity and regularity is the basic skill of human beings. In fact, a person has been learning various rules since birth. For example, what is a rectangle, what is a circle, and what does a table look like, what does a house look like, etc. The high-dimensional image recognition makes the computer have the same image recognition method as humans. When we see that the window of the building in the photo is tilted due to the angle of view, we do n’t think that the window is really tilted. We even know that the window should be square, and we can also tell that it is blocked The branches in front of the window are not part of the window. Similarly, by establishing the law of objects for high-dimensional image recognition, researchers at Microsoft Research have been able to enable PCs to perform similar functions, which can help us correct inclined buildings or wipe off branches in front of buildings.
From this we can also understand the strengths and weaknesses of this technology. Any object or image that follows a certain rule can be identified by establishing a rule. Any object or image that is not regular, this Technology often has greater limitations. For example, processing an object in a chaotic flower cluster is not enough for this technology. Generally speaking, regular objects are often created by human beings, because from the most basic concept of human beings, human beings believe that the world is simple and has a regular structure. When humans create various objects, they will follow simple The principle of easy-to-use, under the influence of this principle, things without regularity will be gradually eliminated. The rules are not as complicated as we think, we do not need to establish a rule for every object in the world. The rule here is actually a classification of mathematical structure. Many objects are the same thing in terms of mathematical structure, so we only need to establish some important passage rules. Of course, there are some special things to establish rules separately, such as text.
The graphic combination invented by human beings, under the adjustment of human aesthetics and regular thinking, has gradually improved and established good regularity, such as horizontal and vertical, and various overall or local vertical and horizontal symmetry. Whether it is English, Chinese or other texts, it basically has a strong regularity, and this kind of rules can also be summarized and summarized mathematically. Such an image has a very low-dimensional internal structure in a high-dimensional space. For example, most Chinese characters have strong regularity except for Chinese characters with few strokes.
In general, this high-dimensional image recognition technology can solve some image recognition problems that we could not solve in the past. After gradually improving, it will completely change the way we recognize and manipulate pictures. High-dimensional image recognition technology Before recognizing the object in the image, the user should tell the computer the location of the regular object being recognized. The next step is to allow the computer to more intelligently discover where there is regularity and where there is no regularity in the image, and use different rules to repair different locations of the image. Another development direction of this technology is to improve computing efficiency, for example, it can eventually be run in real time on terminals such as smartphones.
The auxiliary Transformer for SPS (Switched Power Supply) is a device that converts high voltage AC power from the main transformer into low voltage AC power that is used to power the control circuits and other auxiliary devices in the SPS.
Auxiliary Transformers Of SPS, High frequency transformer, auxiliary transformer, high frequency resonant transformer
Huizhou Show-Grand Electronics Co., Ltd. , https://www.sgtransformer.com