As an electrical engineer, do you know anything about machine vision? Why is it so powerful?

Machine vision is to simulate human visual function through computer, so that the machine can obtain relevant visual information and understand it. It can be divided into "seeing" and "feeling".

"Vision" is to display external information as digital signals through imaging and feed them back to the computer. It needs to rely on a set of hardware solutions, including light source, camera, image acquisition card, visual sensor and so on. Perception is the processing and analysis of digital signals by computers, mainly software algorithms.

Machine vision is widely used in industry, and its core functions include measurement, detection, identification and positioning.

The industrial chain can be divided into upstream parts market, midstream system integration/complete equipment market and downstream application market.

In the upstream of machine vision, there are software and hardware providers such as light source, lens, industrial camera, image acquisition card and image processing software, and in the midstream, there are integrated and complete machine equipment providers, which are widely used in the downstream of the industry. The main downstream markets include electronics manufacturing, automobiles, printing and packaging, tobacco, agriculture, medicine, textiles and transportation.

The global market of machine vision is mainly distributed in North America, Europe, Japan, China and other regions. According to statistics, in 20 14, the global market scale of machine vision systems and components was US$ 3.67 billion, US$ 4.2 billion in 20 15 and US$ 6.2 billion in 20 16. According to the data of North American market, the integration of machine vision system is about six times that of vision system and parts market.

China machine vision began with the introduction of technology in 1980s. With the introduction of 1998 semiconductor factory, it was also included in the machine vision system. Before 2006, domestic machine vision products were mainly concentrated in foreign-funded manufacturing enterprises, and the scale was small. Since 2006, the customer base of industrial machine vision application began to expand to printing, food and other inspection fields, and the market began to grow rapidly at 20 1 1. With the increase of labor cost and the upgrading demand of manufacturing industry, coupled with the rapid development of computer vision technology, more and more machine vision solutions have penetrated into various fields. By 20 16, the market scale of machine vision in China has reached nearly 7 billion yuan.

In machine vision, defect detection function is one of the most widely used functions of machine vision, which mainly detects all kinds of information on the surface of products. In modern industrial automation production, every process of continuous mass production has a certain rate of defective products. Although the proportion is small in isolation, it becomes a bottleneck for enterprises to improve the yield after multiplication, and the cost of eliminating defective products by a complete process will be much higher (for example, if there is positioning deviation in the process of solder paste printing, and this problem is not discovered until online testing after chip mounting, then the repair cost will be more than 100 times of the original cost).

1. Compared with human vision, machine vision has obvious advantages in the inspection industry.

1) high accuracy: human vision is 64 gray levels, and the resolution of small targets is weak; Machine vision can significantly improve the gray level and observe micron-scale targets at the same time;

2) Fast speed: human beings can't see the fast moving target clearly, and the shutter time of the machine can reach microsecond level;

3) High stability: Machine vision has solved a very serious human problem, namely instability. Manual visual inspection is a very boring and hard industry. No matter what kind of reward and punishment system you design, there will be a high rate of missed inspection. The machine vision inspection equipment has no fatigue problem and no mood fluctuation. As long as you write something in the algorithm, you will implement it carefully every time. The controllability of quality control effect is greatly improved.

4) Information integration and retention: The information obtained by machine vision is comprehensive and traceable, and relevant information can be easily integrated and retained.

2. Machine vision technology has developed rapidly in recent years.

1) Image acquisition technology has developed rapidly.

Firmware such as CCD and CMOS are becoming more and more mature, the size of image sensitive devices is shrinking, the number of pixels and data rate are increasing, the resolution and frame rate are changing with each passing day, and the product series is becoming more and more abundant. Parameters such as gain, shutter and signal-to-noise ratio are constantly optimized. The light source, lens and camera are integrated through core test indicators (MTF, distortion, signal-to-noise ratio, brightness, uniformity, color temperature and comprehensive evaluation of system imaging ability, etc.). ).

2) Image processing and pattern recognition are developing rapidly.

In image processing, with the extraction of high-precision edge information, many low-contrast defects that were originally mixed with background noise and difficult to detect directly began to be distinguished.

Pattern recognition itself can be regarded as a marking process, which divides the pattern to be recognized into its own patterns on the basis of certain measurement or observation. Decision theory and structural method are widely used in image recognition. The basis of decision theory and method is decision function, which classifies and recognizes pattern vectors based on time series description (such as statistical texture). The core of structured method is to decompose an object into patterns or pattern primitives, and different object structures have different primitive strings (or strings). Using the given pattern primitive to find the coding boundary of the unknown object, the string is obtained, and then its genus is judged according to the string. In the aspect of feature generation, many new algorithms appear constantly, including features based on wavelet, wavelet packet, fractal and unique binary component analysis. There are also tube support vector machine, deformable template matching, linear and nonlinear classifier design and so on.

3) Breakthrough brought by deep learning

Traditional machine learning mainly relies on people to analyze and construct logic in feature extraction, while deep learning simulates brain work through multi-layer perceptrons, and builds deep neural networks (such as convolutional neural networks) to learn simple features, build complex features, learn mapping and output, and all levels will be continuously optimized in the training process. In specific applications, such as automatic ROI region segmentation; Punctuation position (unknown defects can be detected flexibly through error-proof vision); Re-detect defects that cannot be described or quantified from high-noise images, such as orange peel defects; Distinguish the true and false defects in the detection of glass cover plate. With more and more machine vision software based on deep learning coming to market (including vidi in Switzerland, SUALAB in South Korea, ASTRI in Hong Kong, etc.). ), the empowerment of deep learning to machine vision will become more and more obvious.

4) The development of 4)3d vision

Three-dimensional vision is still in its infancy, and many applications are using three-dimensional surface reconstruction, including navigation, industrial inspection, reverse engineering, surveying and mapping, object recognition, measurement and classification. However, the accuracy problem limits the application of 3D vision in many scenes. At present, the first application in engineering is the volume measurement of standard parts in logistics, which is believed to have great potential in the future.

3. There are still many difficulties to be overcome before machine vision can completely replace artificial vision detection.

1) light source and imaging: high-quality imaging in machine vision is the first step. Because the surface reflection and refraction of different materials will affect the feature extraction of the measured object, light source and imaging can be said to be the first difficulty to be overcome in machine vision inspection. For example, at present, the scratch detection of glass and reflective surface is often stuck in the integrated imaging of different defects.

2) Feature extraction in high-noise and low-contrast images: In high-noise environment, it is often difficult to identify true and false defects, which is why there is always a certain false detection rate in many scenes. But with the rapid development of imaging and edge feature extraction, this area has been making various breakthroughs.

3) Identification of unexpected defects: In application, some specific defect patterns are usually given, and machine vision is used to identify whether they have occurred. However, many obvious defects are often missed because they have never happened before or in too many ways. If you are a person, although he is not required to detect this defect in the work flow file, he will notice it, so there is a greater chance to catch it, and the "wisdom" of machine vision in this respect is still difficult to break through at present.

4. Machine vision industry chain situation

1) Upstream parts market

Mainly including light sources, lenses, industrial cameras, image acquisition cards, image processing software and other providers, in recent years, smart cameras, industrial cameras, light sources and boards have maintained a growth rate of not less than 20%. According to the survey of China Machine Vision Industry Alliance (CMVU), there are nearly 200 international machine vision brands that have entered China at present (such as core component manufacturers represented by cognex, Darsa and Baomeng, and those represented by Kearns, Omron, Bonner and Ni are also involved in machine vision core components and system integration). There are more than 65,438+000 domestic machine vision brands (such as Haikang, Huarui, Monto Optoelectronics, Shenzhou Vision, Shenzhen Canrui, Shanghai Fangcheng, Shanghai Bochuang Electric, etc.). ), and there are more than 300 machine vision agents (such as Shenzhen Hongfu Vision, Microvision New Era, Sanbao Xingye, Ling Guangyun, Sunshine Vision, etc.). Many domestic machine vision parts markets started from acting as agents for foreign brands, and many enterprises have good cooperation with their foreign counterparts, and this cooperation is exclusive, which brings a certain threshold to potential entrants, so agents of high-quality products also have good market competitiveness and profit performance. At the same time, the core components of domestic industrial vision represented by Haikang and Huarui are rising rapidly.

2) Midstream system integration and whole machine equipment market

There are more than 100 system integration and machine equipment suppliers in the middle reaches, which can provide comprehensive machine vision solutions for automation companies in various industries, such as Ling Guangyun, Microvision New Era, Jiaheng, Sunshine Vision, Daheng Image, etc. Because there is still a big gap between domestic products and international products, and many midstream system integrators and complete equipment suppliers start from the trade of core components, many of them prefer foreign brands in the choice of visual products. In order to promote their own software and hardware products, domestic brands often need to develop their own solution integration capabilities in order to better face market competition.

3) Downstream application market

The downstream of machine vision is mainly a company that provides non-standard automation integration solutions for end users. Industry attributes are very strong, and the core competitiveness is a comprehensive understanding of industry and production and the integration of various technologies. Because the change of industry automation has a certain periodicity, it is deeply influenced by the overall upgrading speed, shipment volume and profit of the industry. Therefore, the most important thing to promote the popularization of machine vision applications in the past two years is the electronics manufacturing industry, followed by automobiles and pharmaceuticals.

1. Semiconductor and electronic manufacturing industry: From the application distribution of domestic machine vision industry, 46% are concentrated in electronic and semiconductor manufacturing industry, including wafer processing and manufacturing, PCB inspection (negative film, inner/outer board, final appearance inspection of finished products, etc.). ), SMT mounting inspection, AOI defect inspection of LCD whole process, surface defect inspection of various 3c components, appearance inspection of 3c products, etc.

Two. Automobile: body assembly inspection, geometric dimension and error measurement of parts, surface and internal defects inspection, clearance inspection, etc.

Three. Printing and packaging inspection: tobacco shell printing, food packaging printing, aluminum-plastic board packaging printing, medicines, etc.

Four. Agriculture: classification, inspection and grading of agricultural products

Verb (abbreviation of verb) textile: detection of foreign fiber, cloud weaving, warp and weft defects, identification of fabric surface fluff, yarn structure analysis, etc.

5. Future development trend of machine vision system

1) Embedded solutions are developing rapidly, and smart cameras has outstanding performance and cost advantages, so embedded PC will become more and more powerful.

2) Modular general software platform and artificial intelligence software platform will reduce the technical requirements for developers and shorten the development cycle.

3)3d vision will move towards more application scenarios.