Machine vision inspection systems are widely used in industrial production due to their high precision and non-contact operation, including workpiece positioning, measurement, and identification. Establishing a stable and reliable vision inspection system is the goal of vision system designers. This paper first introduces the basic components of a machine vision inspection system, including sensors, light sources, and inspection software. It then analyzes the factors that lead to instability in vision systems and finally proposes corresponding solutions and suggestions to address these factors affecting stability.
1. Introduction
The study of machine vision began in the 1950s with pattern recognition of two-dimensional images [1]. It was initially designed to replace the human eye in detection and recognition work, which can greatly improve the efficiency of detection work and reduce the inconsistency of detection results caused by human eye fatigue. Machine vision detection has developed to the point that it can accomplish many tasks that are difficult for the human eye to complete, such as high-precision measurement and high-speed classification of specific products, as well as using infrared, ultraviolet, X-ray and other detection technologies to detect things that cannot be detected by human vision [2]. However, the difficulty in designing machine vision systems lies in how to ensure their reliability and stability. Whether from the design of hardware such as light source and camera or from the design of image processing software, it has an important impact on the stability of machine vision.
2. Composition of a machine vision system
A typical machine vision system generally consists of three parts: image acquisition, image processing and analysis, and output or display. Based on their structural composition, vision systems are mainly divided into two categories: PC-based or board-based machine vision systems, and embedded machine vision systems, also known as "smart cameras".
2.1 PC-Based Vision System
A PC-based vision system is a vision system based on a personal computer (PC). Its image acquisition equipment generally consists of a light source, an optical lens, a CCD or CMOS camera, and an image acquisition card. The image processing and analysis equipment is based on a PC and uses image processing software. The display output of the image processing results is generally a monitor.
PC-based vision systems have evolved to meet diverse needs in various working environments. Camera options range from 2 megapixels to 12 megapixels, and frame rates from 0 to hundreds of frames per second or even higher. Communication is highly flexible, supporting direct use of USB (USB 2.0/USB 3.0) interfaces, Gigabit Ethernet (GigE) interfaces, or easy expansion with cameras using FireWire (1394a/1394b) and CameraLink interfaces. They also exhibit strong anti-interference capabilities at short distances. In terms of speed and accuracy, PC-based systems can be configured with high-speed, high-resolution cameras and high-speed processors to meet the demands of high-speed motion or high-precision detection.
However, PC-based machine vision application systems are relatively large, requiring not only cameras but also image acquisition cards, industrial computers, and various connecting cables. This makes them unsuitable for applications with strict size constraints, such as within production facilities or on transfer devices. Furthermore, their complex structure is widely considered to involve numerous external components manufactured by different companies, leading to compatibility issues and intermediary connections, resulting in lower integration and decreased stability. Compared to highly integrated smart cameras, their development cycle is also relatively longer.
2.2 Embedded Vision System
A smart camera mainly consists of three parts: an image acquisition unit, a communication module, and an image processing unit (processing software). The image acquisition unit is similar to a traditional camera, converting light signals into analog or digital signals; it's equivalent to a CCD/CMOS camera and image acquisition card. The image processing unit is like the PC part of a PC-based system, the core of embedded vision, including image processing, storage units, and corresponding processing software. Generally, three hardware platforms—DSP, FPGA, and RISC—are used to perform image processing operations. The software can be written externally, but mature embedded machine vision systems typically encapsulate common image processing algorithms into fixed modules, which developers can choose to call as needed. The communication module is also an important component of a smart camera, mainly outputting the image processing results externally. Smart cameras generally have a built-in Ethernet communication module and support various network and bus protocols.
Compared to PC-based vision systems, embedded vision systems offer several significant advantages. For instance, because their image acquisition and processing units are directly connected, they exhibit better pixel consistency and stronger anti-interference capabilities. Furthermore, due to the high level of integration, smart cameras are much smaller than PC-based vision systems, making them suitable for harsher working environments. Because of their high integration, the image acquisition, processing, and communication components of smart cameras undergo rigorous reliability testing during design and manufacturing, resulting in significantly higher operational stability than PC-based systems. Their simpler structure also simplifies maintenance. On the software side, mature smart cameras have essentially pre-built vision algorithm and communication modules, requiring only simple user calls. Unlike PC-based systems, they do not require the complex, low-level development required for PC-based systems, simplifying development and shortening the development cycle.
However, embedded vision systems also have certain disadvantages compared to PC-based vision systems. For example, in terms of accuracy and speed, due to limitations in size and image processing capabilities, smart cameras cannot easily integrate high-speed or high-resolution cameras into vision systems like PC-based systems. Under current technological conditions, achieving the same resolution and speed with smart cameras typically requires higher costs due to more demanding manufacturing processes and circuit design issues. Furthermore, the highly integrated nature of smart cameras also leads to a lack of flexibility. Their hardware and software development is relatively fixed, and their scalability is worse than that of PC-based systems. In complex machine vision scenarios, sometimes it is difficult to complete the system's functional design using only a smart camera, while PC-based systems can select suitable and inexpensive equipment based on the actual situation, and can also choose different third-party software to complete tasks such as image processing.
3. Instability factors of machine vision inspection systems
3.1 Introduction to Imaging Systems and Instability Factors
The imaging system mainly consists of a camera (CCD/CMOS), a lens, and a light source. It is the foundation of visual inspection. The purpose of the imaging system is to acquire qualified raw images. A good imaging system must ensure the stability of image quality during system operation. Stable image acquisition is the basic guarantee of the stability of visual inspection.
3.3.1 The impact of industrial cameras on imaging stability
For vision system designers, the selection of industrial cameras mainly considers their sensor type, resolution, and frame rate. Sensors are divided into two types: CCD and CMOS. CMOS image sensors have high integration, and the distance between various components and circuits is very close, resulting in more severe interference and high imaging noise. Compared with CMOS cameras, CCD sensor cameras have the characteristics of high sensitivity, low noise, and fast response speed. In terms of stability, CCD cameras are also more resistant to shock and vibration. Generally speaking, CCD sensor cameras are superior to CCD cameras in terms of image quality and stability.
Another important factor affecting camera image quality is the camera lens. In addition to selecting appropriate parameters such as focal length, depth of field, and aperture according to specific working conditions, a significant factor affecting the system's detection accuracy is geometric distortion error. This is an inherent transmission distortion of optical lenses, which is affected by the manufacturing process and cannot be eliminated, only compensated for. Although many industrial cameras now use various methods to compensate for errors caused by lens distortion, geometric distortion will still affect detection accuracy in high-precision detection fields.
3.1.2 Influence of light source on imaging stability
Light sources have the function of magnifying the features and defects of images, reducing clutter and background, and directly affecting the quality of input data. Since there is no universal lighting equipment, the design of light sources has always been a difficult point in machine vision systems. Usually, it is necessary not only to select the type of light source for each specific application instance, but also to consider the installation of the light source and the illumination method of the light source according to the specific environment in order to achieve the best effect. Different types of light sources have different stability. Common visible light sources include LED light sources, halogen lamps, fluorescent lamps and sodium lamps. The biggest drawback of visible light is that it cannot output light energy continuously and stably. For example, the light energy of a fluorescent lamp will decrease by about 15% in the first 100 hours of use [4]. As the usage time increases, the light energy output continues to decrease. In addition to visible light, in high detection task scenarios, invisible light such as X-rays and ultrasound are often used as light sources. They can output light energy continuously and stably, but they are not conducive to the operation of the detection system and are expensive. The non-uniformity of the light source will also affect the image quality. The difference in luminous intensity in different directions will also cause noise. LED light sources in the visible light spectrum offer better stability and lifespan compared to halogen lamps and fluorescent lamps, with shorter response times, customizable colors, and lower operating costs, leading to their widespread application. Illumination methods for these light sources can be categorized as backlighting, front lighting, structured light lighting, and strobe lighting, with the design principle being to highlight image features.
a. Fluorescent lamp light source b. LED light source
3.2 Software Stability
The impact of detection software stability on machine vision is undeniable. The vision system will ultimately use software on a computer to perform a series of image processing tasks such as image filtering, edge detection, and edge extraction. Different image processing and analysis methods, as well as different detection methods and calculation formulas, will bring different errors. The quality of the algorithm determines the level of measurement accuracy.
3.3 Environmental Factors
The measurement environment for vision systems includes temperature, light, power supply changes, dust, humidity, and electromagnetic interference [5]. A good environment is essential for the normal operation of vision systems. External light affects the total light intensity illuminating the object being measured, increasing the noise in the image data output. Changes in power supply voltage can also cause instability in the light source, generating noise that changes over time. Temperature changes can also affect the performance of the camera. Cameras are marked with their normal operating temperature range when they leave the factory. Overheating or overcooling can affect the normal operation of the camera. Electromagnetic interference is an unavoidable interference factor in industrial testing sites. It has a particularly serious impact on weak electrical circuits such as industrial camera circuits and data signal transmission circuits. Qualified vision products undergo rigorous anti-interference testing when they leave the factory, which greatly reduces the impact of external electromagnetic interference on hardware circuits.
3.4 Influence of Mechanical Structure Positioning
Besides the imaging system hardware, the relative positional relationship between the camera and the object also affects the stability of image quality. For example, vibrations in the mechanical support structure of the camera or workpiece can affect detection accuracy, and this is a difficult problem to troubleshoot. When inspecting a workpiece dynamically, the impact of motion blur on image accuracy needs to be considered (blurred pixels = object speed * camera exposure time). Furthermore, ideally, the optical axis of the CCD camera lens should be perpendicular to the plane of the workpiece. However, in actual use, due to installation errors or manufacturing errors of the camera or workpiece, the optical axis cannot be guaranteed to be perfectly perpendicular to the measured plane, resulting in a certain angular deviation, which also affects measurement accuracy.
4. Solutions to ensure stability
4.1 Hardware Selection and Design
The choice of imaging system hardware is particularly important. As can be seen from the above analysis of CCD and CMOS cameras, unless there are special requirements, such as a high shooting speed (CMOS has a faster readout speed), CCD sensor cameras are the primary choice to ensure image quality and stability. The resolution and frame rate of the camera are mainly selected based on the detection accuracy and detection speed. The appropriate resolution is determined by calculating the field of view of the object to be detected and the distance between the camera and the object. The frame rate of the camera is selected by considering the movement speed of the object and the detection accuracy requirements.
For lenses, the primary consideration is selecting a lens resolution that matches the camera's maximum resolution. Choosing a lens with a resolution greater than the camera's maximum resolution is sufficient. The focal length of the lens also needs to be calculated based on the working distance and field of view, and an appropriate depth of field should be selected according to the changes in distance between the object being measured and the camera. For high-precision measurements, to ensure accuracy, in addition to correctly selecting the above parameters, a telecentric lens with lower geometric distortion compared to ordinary lenses can be chosen. Telecentric lenses not only have lower geometric distortion but also reduce errors caused by changes in object distance.
Unless otherwise specified, invisible light sources such as X-rays are used. For visible light sources, LED light sources should be given priority. LED light sources are significantly superior to halogen lamps, fluorescent lamps, and other light sources in terms of light source uniformity, which has a decisive impact on the quality of acquired images. Furthermore, LED light sources have advantages such as low power consumption, long lifespan, and no environmental pollution. Meanwhile, to reduce the impact of external light on the stability of the vision system, external light sources can be shielded by adding a light source enclosure.
4.2 Software Design
4.2.1 Calibration
Due to manufacturing processes, cameras and lenses inevitably produce geometric distortion errors in the acquired raw images to varying degrees. These errors cannot be eliminated through hardware optimization, but their impact on measurement accuracy can be mitigated using calibration software algorithms. The basic principle of camera calibration is to determine the camera's intrinsic and extrinsic parameters, as well as distortion parameters, by taking pictures of standard images (usually using a calibration board) at different angles within the field of view. This establishes a mapping relationship between three-dimensional coordinates and image coordinates, thereby correcting the obtained raw distorted image. Camera calibration is typically essential in measurements and positioning where high precision is required.
4.2.2 Image Processing
The raw images acquired by the hardware ultimately need to undergo image filtering, edge detection, and other algorithms to complete the detection function and output the detection results. Image filtering suppresses noise in the acquired images, reduces instability in light sources and grayscale values, and improves the signal-to-noise ratio. Essentially, it uses algorithms to ensure the minimum variance between pixels in the image. For high-precision measurement systems, coarse-grained pixel-level accuracy is often insufficient. Subpixel-level edge localization technology, through a combination of subdivision algorithms and fitting methods, can achieve subpixel-level accuracy of 0.1 or even 0.01 at the pixel level, ensuring the system's detection accuracy.
5. Summary
In summary, the design of a machine vision system requires consideration of multiple factors. Besides selecting equipment with appropriate parameters based on standard requirements, it's also necessary to consider the stability of the light source, camera distortion errors, and the stability interference and measurement errors caused by the relative motion between the object being inspected and the camera. Only by comprehensively considering these factors and optimizing the vision system design can a stable and qualified machine vision inspection system be established.
6 Industry Leaders
WELINKIRT is a privately held machine vision technology company. It possesses leading vision-guided robot (VGR) software products, a platform that enables industrial and collaborative robots to "see and think."
Artificial intelligence and machine learning are areas of expertise for WELINKIRT.
The difference between intelligent robots and ordinary robots lies in their ability to adapt to constantly changing environments. Vision, as the robot's "eyes," is crucial. Robots require sensory input from their operating environment, and vision is the primary sensory input.
Werobotics provides contextual intelligence for robots.
WELINKIRT's proprietary algorithm is already a form of supervised artificial intelligence, allowing the robot to "think." With the introduction of machine learning techniques into Werobotics, we will create unattended vision software that trains itself faster and more accurately than humans.
Core Product
WELINKIRT leads in the three main areas of robot vision (2D, 2.5D, and 3D), providing vision robot solutions and owning the award-winning Werobotics software platform.
Single-camera 2D, single-camera 2.5D, single-camera 3D
Product Features
Hardware: Simple structure, quick deployment, and inexpensive finished product.
The software has three main features: speed, convenience, and automation.
AutoCalibration automatically measures and calibrates the position information of the camera and robot.
AutoTrain automatically moves the 3D model of the component object around the workpiece.
AccuTest automatically tests and simulates to verify the accuracy of the results after the parameters are set.
Product Capabilities
1. Built-in verification tool
2. Automatic Vision System Calibration Test - AccuTest
3. Vision-Guided Robot (VGR)
4. Random sorting
5. Automatic calibration
Application scenarios
Erobotics Technologies' vision technologies are applicable to various industries, types of applications, and environments. Some applications are as follows:
1. Structured picking
2. Semi-structured picking
3. Completely unstructured random selection
4. Inventory Organization and Management
5. Motion Detection
6. Conveyor belt workpiece tracking and handling
7. Automotive parts processing and assembly
Microchain Cognitive Robot Vision System
Less hardware investment, higher performance
A single lens can satisfy 2D, 2.5D, and 3D visual applications.
Fastest: 0.1 seconds from image to boot
Highest precision: 20 micrometers, or 0.02 mm
Lowest price: Reduce supplier costs by 40-50%
No programming required, and it's easy to use.
Let machines understand the world like humans do.
