Machine Vision vs Computer vision construction of knowledge system

the difference and connection between machine vision and computer vision

In many cases, we mistakenly believe that machine vision is the computer vision, in fact, this is not accurate. What is machine vision. What is computer vision. First, we start with the definition, machine vision is to use machines instead of human eyes for measurement and judgment. Computer vision is the use of computer and its auxiliary equipment to simulate human visual function, to achieve the objective world of three-dimensional scene perception, recognition and understanding. Machine vision and computer vision are not only two different concepts, but also different emphases. Machine Vision focuses on the application of engineering, emphasis on real-time, high-precision and high speed, while computer vision focused on theoretical algorithms, emphasizing the theory, because the theoretical research and development speed is often faster than practical applications, that is, the development of computer vision is far more than the speed of its time production application speed, Therefore, many computer vision technology is still difficult to apply to machine vision. But the two still share a set of theoretical system, but the direction of development is different, a focus on practical applications, a focus on theoretical algorithm research, can not say who replaces who, different.

Second, the situation of visual development in recent years and the bottlenecks encountered

1, the algorithm bottleneck. the research object of machine vision is mainly image and video, the images and videos we collect are characterized by large data, redundant information, high feature space dimension, and considering the diversity of objects and problems faced by the real machine vision, single simple feature extraction algorithm (such as color, space orientation and frequency, Boundary shape and so on) is difficult to meet the requirements of the algorithm for general suitability, so the design of the universal feature extraction algorithm for the computational capacity and storage speed is very large, which resulted in a significant increase in the cost of development.

2, scene cognition problem. How to make machines recognize the world. There is no mature answer to this question at present, and it is the hot direction that scientists have been studying all the time. The early development of artificial intelligence has undergone a series of development of Semiotic School, behaviorism School and connecting school, but no satisfactory answer has been found, at present newer ideas think that we should construct intelligent machine vision system from analyzing, understanding and simulating the information processing function of human brain. But the development of neuroscience can only be done to understand and simulate a part of the brain, rather than the whole (of course, computational capacity constraints are also one of the reasons). In fact, the question of how people perceive a goal or scenario remains in the qualitative description rather than the quantitative description .

3, the accuracy of the problem. one of the problems that machine vision systems are often criticized for is accuracy. For example, the face recognition algorithm, which was in full swing 10 years ago, although a series of seemingly excellent algorithms are available, the algorithms are currently in the specified sample library, while the accuracy of face recognition is still not enough to meet the needs of the actual application in an unspecified large sample library Therefore, it is impossible to replace the fingerprint or iris close contact Biometric identification method. It is no coincidence that the problem arose. Because the more sophisticated the goal, the more complex the information, the greater the ambiguity and uncertainty. Human face can be better to identify, in fact, at the expense of a certain accuracy. While machine vision is doing things to learn from the human brain or human eye system inspiration to deal with the complex and huge flow of information, on the other hand to eliminate the human brain in the pattern recognition of the lack of precision deficiencies. This is clearly a wishful thinking approach.

4, robustness problem. compared with other measuring means, the biggest advantage of vision is that it can quickly obtain three-dimensional information, one or several photos can reconstruct the measured object's three-dimensional characteristics, and then achieve measurement. However, as we know, as long as the measurement conditions, the environment, the surface characteristics of the measured material changes, sometimes even slightly changed, the result is very different, the measurement of repeatability and accuracy is more impossible to talk about. This is also the current machine vision measurement dimensions, posture and other parameters of the outstanding problems, especially in some strong light interference, temperature field changes, lighting conditions of the application of the situation is particularly prominent.

5, the lack of talent . At present, the practitioners in the real sense lack of trained, lack of the underlying theory of image processing cognition and understanding. Machine Vision image processing is a very important link, and most of the current practitioners are undergraduate or college graduate, or electrical engineers to the new entry, the basic lack of the basic theory of image processing. While compared to the average automation practitioner, Machine Vision engineer treatment is good, but it is difficult to attract a master's or PhD in specialized image processing academic training to join, because casually join that big Internet company to do image-related work, treatment can be automated engineers out of several streets. In addition, machine vision more applications belong to automation equipment this piece. And automation is a more interdisciplinary subject, involving machine vision, need to understand things including, electrical, motion control, machinery, optics, software programming and so on. It is not difficult to understand some basic things in these subjects, but it is more difficult to study them thoroughly and to use them efficiently.

The current industrial vision applications include: detection, measurement, identification and positioning. And these few aspects of machine vision is not a real sense of the implementation of mass detection at the same time to ensure a very high accuracy rate, the minimum rate of false detection and eliminate. This goal can not be achieved, reducing the application of machine vision expectations. Because machine vision equipment can not be completely solved, still need a person to review, unless the customer's standard is not so high. This also led to the current machine vision in the industrial application is not so rapid popularization of one of the reasons. Why machine vision is encountering bottlenecks. The main is not past the customer that a close--high precision, high speed, high accuracy, and real-time is also better.

Domestic hardware core components (cameras and lenses) and software algorithm package or a foreign product leader, domestic also appeared a number of alternative products, from the performance and the foreigner PK still have a big gap.

Reference Documents

1, http://www.zhihu.com/question/20023867