Benefits of High-End Machine Vision Systems

In the ‘‘old days’‘ virtually every product in the machine vision world was a ‘‘high-end’‘ system. The challenge was that with limited computing power they could do very little. If, in fact, one attempted to enable the ‘‘high-end’‘ functions, they would inevitably wind up being too slow for the application. Hence, many of the early gray scale systems quickly reduced their image to a binary representation of the scene, generally by using simple fixed threshold techniques. This resulted in marginal performance at best.

Today, high-end systems capable of performing any number of robust preprocessing and segmentation algorithms without such processing becoming the bottleneck are commercially available. One might say these vision processors are truly ‘‘general-purpose’‘ machine vision systems, as they permit one to fine tune processing for a specific application. These systems perform multiple preprocessing and segmentation algorithms at real-time rates using as the basis of their designs: application-specific integrated circuits (proprietary or commercially available), FPGAs, DSPs, microprocessors and/or combinations of these compute capabilities.

These systems are distinct from those referred to as smart cameras, embedded vision computers/processors or host-based processing systems that employ simple frame grabbers or digital cameras that deliver image data to the host directly. While these systems do have a general-purpose flavor to them, in general one has access to a very limited suite of the underlying preprocessing and segmentation algorithms. The machine vision software purposefully makes the underlying image processing technology transparent to the user. Interface is via icons representing specific machine vision functions (gauging, finding, flaw inspection, optical character verification, etc.). There is limited opportunity to get ‘‘under the hood,’‘ so to speak, to tweak parameters.

These ‘‘lower-end’‘ machine vision products can handle many applications today, as appearance and presentation variables encountered in many applications are normalized with the latest in commercially available dark field and bright field lighting arrangements and cameras with features such as progressive scan and asynchronous reset. However, where such appearance and positional variables cannot be normalized and/or where throughput requirements are high, these ‘‘lower-end’‘ products will not result in rigorous performance. Hence, there are still opportunities for high-end machine vision systems.

To gain insights for this article a questionnaire was sent to companies known to be providers of high-end machine vision systems. Responding to the questionnaire were:

George Blackwell – Cognex Corporation
Stan Karandanis – Datacube
Yves Joskin – Euresys
John Marioni – Newton Labs
Joe Germann, SKY Computers®

1. What exactly constitutes a ‘‘high-end machine vision system’‘ these days?

John Marioni: ‘‘Any one or more of the following:

• $20,000+ price tag
• Cannot be done with a smart camera
• Involves tricky camera control and or light techniques
• Involves multi cameras
• Requires heavy processing
• Requires special communications
• Requires special image display
• Requires special software algorithms’‘

Yves Joskin: ‘‘Probably the best way to define a high-end vision system is to define beforehand what a mainstream system is. Today, a mainstream system is based on an off-the-shelf desktop-type PC equipped with a frame grabber specialized in acquisition of images issued by an industrial camera. The image processing tasks of mainstream systems rely on standard imaging libraries. An engineer reasonably skilled in software code writing will develop the machine vision application. The PC runs a standard operating system, namely Microsoft Windows or Linux. When one of the characteristics of the machine vision system outperforms the capabilities of such a mainstream system, one is faced with a high-end system.’‘

Stan Karandanis: ‘‘What was high 5 years ago is run of the mill and available for under $1K.The question that should be asked is ‘‘can the job be host done?’‘  If the answer is yes the only contribution the frame grabber or processor does is transfers the signal to the host, something that Firewire or USB2 does quite nicely.  If the host is not capable of providing a timely solution then there is a chance that some sort of interface accelerator will find a home.’‘

Joe Germann: ‘‘High-end machine vision systems process real time, high volume, continuous data streams from optical or specialized sensor sources with typical data rates in the 10's to 100's of MB/sec of continuous real time data This data may not actually begin as image data but rather as data that requires complex and demanding image reconstruction algorithmic processing before the actual image processing begins Processing demands are typically in the order of 100 to 1000 CPU operations per image pixel.  These high-end systems are typically used to detect the presence or lack thereof of a specific image attribute. i.e., looking for tanks under trees.’‘

George Blackwell: ‘‘As a full-line vision supplier, Cognex Corporation considers high-end systems those that are programmable, PC-based systems utilizing the most sophisticated algorithms for flexibility, accuracy, and robustness.  High end doesn't necessarily mean high price. If there's already a PC used for machine control, many times a PC-based system can cost less than a vision sensor.’‘

2. What are the features of such systems?

Germann: ‘‘A modern high-end machine vision system includes:

• Open systems architecture must be designed to ensure high application availability and quick application deployment.  These systems must be faster, less expensive and deployed with minimal risk.  Pressures to reduce maintenance costs require that systems be built with less costly and more reliable components, including both hardware and software. Critical, harsh environments require that systems be reliable and resilient, with uptimes equaling those found in business-critical corporate data centers.

   

• A Data acquisition server that is interfaced to the sensor to manage the data flow from data acquisition, to data redistribution to the image-processing engine.
• Scalable architecture to easily expand the system to handle the forever increasing image stream data rates and image processing algorithmic complexity.
• Systems must employ ‘‘smart’‘ systems architectures. A smart architecture guarantees data communications integrity and adapts to system changes caused by hardware failures, software errors or resource additions.  It enables remote management by integrating hardware and software technologies that provide system sensing, monitoring, and proactive and reactive management capabilities. With multi-chassis scalability, the architecture adapts to resource additions. A smart architecture provides proactive failure management and assures high application availability. It responds to the needs for resiliency, security, remote manageability, graceful degradation, scalability, portability, adaptability and quick and easy application development and implementation. 
• A high-end image processing system, especially when continuous image streams are part of the problem set, are best constructed by taking into account the natural decomposition of the problem.  In particular there is a definitive piece that is responsible for the data acquisition, buffering, and management of the incoming image streams.  Likewise there is a piece of the solution that focuses on the actual image reconstruction and processing at the algorithmic level.’‘

Blackwell: ‘‘High-end systems come with a full library of vision tools capable of handling the most complex applications. They are programmable, provide users with access to more parameters, and typically capable of managing large databases.  Since high-end systems are PC based, their performance increases with each boost in PC processor speed. ‘‘

Joskin: ‘‘A system is called ‘‘high-end’‘ when one or several of the following requirements come true.

• The raw data throughput of the data flow issued by the camera, and conditioned by the frame grabber, exceeds the bandwidth of the mainstream PCI bus.
• The above limit is reached because several cameras are operated in parallel.
• The reactivity of the system cannot be achieved with the latencies associated with the mainstream operating systems.
• The algorithms used by the application cannot be found in a standard machine vision software library.
• The targeted mechanical robustness is not met with a low-cost desktop PC.

The mainstream PCI bus referred to in this discussion features 32-bit and 33 MHz.’‘

Marioni provided the following table:

3. How do you differentiate ‘‘high-end systems’‘ from ‘‘low-end systems’‘? What does one find in ‘‘high-end systems’‘ that you don't find in ‘‘low-end systems’‘?

Blackwell: ‘‘The key differentiator is the development environment. Low-end systems tend to have a more ‘‘configurable’‘ environment while high-end systems have a ‘‘programmable’‘ environment. In our case, for example, with the In-Sightâ machine vision sensors, users configure a vision application within a spreadsheet using menus and dialog boxes. There is no programming language involved.  Our high-end systems, in contrast, allow users to choose among different software development environments based on individual skill and experience levels. This enables users to build powerful vision applications using the development environment they are most comfortable with.’‘

Joskin: ‘‘High bandwidth systems: A vision system involving exceptionally high throughput will use a specific strategy. The simplest strategy consists of enabling the frame grabber with a non-mainstream interconnect bus, such as the 64-bit, 66 MHz PCI bus or the highest performance PCI-X. Another trick is to accumulate the high-speed camera data into a large on-board memory, and to deliver it to the PC at a compatible rate. In this case, the high-end system differentiates from the low-end because of this on-board large-size buffer. A more complex strategy involves some means to reduce the incoming data flow to a bus-acceptable amount. Such a high-end frame grabber is sometimes called a frame processor. Several types of processing devices are found, ranging from very specialized FPGAs to general purpose CPUs. The general idea of preprocessing is to perform some kind of feature-extraction process resulting in a data set holding the information essential for the application, but with a considerably reduced byte count.

High reactivity systems: System architects may find it inadequate to rely on PC software to handle the time-sensitive part of the machine vision application. The argument for this is usually that the standard operating system running on the PC is not designed for low latency and deterministic real-time operation. A safe and predictable operation would be achieved only if the compute intensive vision analysis tasks are decoupled from the user-oriented tasks, such as interacting with the operator and building reports. The defenders of this opinion are satisfied by a high-end acquisition board with on-board vision preprocessing or processing capabilities, which is viewed as a vision subsystem with a distinct role. In particular, when very specific or confidential proprietary algorithms are needed, this kind of high-end board is a viable solution.

High dependability systems: The mechanical performance of a standard PC may be considered as not satisfactory for certain classes of highly professional applications, such as medical or military. In this case, the high-end nature of the system will be materialized by a different mechanical form-factor, such as Compact PCI. Environmentally resistant enclosures are more easily found with these types of industrial PCs.’‘

Marioni: ‘‘High End: the vision system is catered to the specific and unique customer requirements, meeting 100% of specifications an expectations. Low End: the given vision system (which is inherently limited in speed, processing, features, etc.) is applied as a ‘‘best fit’‘ for the application, but is not necessarily optimal for the task nor capable of meeting 100% of objectives.’‘

Germann: ‘‘Low-end systems are typically found in inspection applications such as examining a PCB for solder shorts or on repetitive commodity production lines. Typically connected to a camera interface, a low-end PC-based system can usually meet the data acquisition and computational requirements for these machine vision applications. High end systems include:  radar applications such as synthetic aperture radar, foliage penetrating radar, real time cardiac imaging, signal intelligence, and explosive detection in luggage.  All of the high-end applications use complex algorithms and high-speed computer processing to examine fine details within the huge data streams.  Typically connected to data acquisition that is either optical or contains many sensors, these are large multi- to many- processor systems and subsystems.’‘

4. How are they physically different?

Joskin: ‘‘A high-end frame processor is larger and more expensive than a low-end grabber. This is obviously the result of the additional components implementing processing devices and additional memory. The high-end board may exhibit an unusual form factor and/or PCI interface.’‘

Germann: ‘‘PC vs. large rack, chassis or cabinet sized computer.’‘

Marioni: ‘‘Physical difference between high and low end systems: 

• High end systems have control boxes (to house the processors and IO) to which the cameras attach 
• Low end systems contain both camera and processor in one housed unit.
• High end systems contain a specific sub set of software and avoid unnecessary software overhead
• Low end systems contain a general purpose software toolkit most of which is not necessary for a given application, nor necessarily optimum for the inspection task.
• High end systems have much shorter software manuals
• Low end systems have huge software manuals to cover the broader toolkit overkill.’‘

5. What are the characteristics of some applications where the properties of a high-end system are critical to the success of the application?

Germann: ‘‘The characteristics of applications which require high-end systems are: 

• Real time, continuous data stream
• Real time analysis requirement
• Complex algorithms which are not the ‘‘typical’‘ image processing algorithms, require very powerful computing to solve in a timely (real time) manner
• High application availability of multi- to many- processor systems. A high-end Imaging system can have 100’s of image processing nodes working in parallel with each other
• Quick application implementation of multi- to many- processor systems’‘

Joskin: ‘‘Typically, high-end systems come in to play when the raw sustained image data bandwidth exceeds 100 MB/s. Once the sustainable bandwidth rate of the fastest PCI-X bus is exceeded (700-800 MB/s), then it is absolutely necessary to have a high-end system with embedded processing capabilities. For in-between situations, a high-end device will usually use a high-end PCI bus.  Life-critical applications may ask for an embedded frame processor, running a specialized real-time operating system (RTOS). Severe environmental conditions (e.g. humidity, dust, vibration…) will call for a high-end frame grabber exhibiting an industrial form factor.’‘

Blackwell: ‘‘High-end systems are generally used for more mathematically intensive applications, such as defect detection that require sort and classify algorithms, or when accuracy requirements are up to 1/40th of a pixel. Even though a machine vision sensor may be able to accomplish the task when speed requirements exceed 1500 parts per minute, a high-end system would provide additional headroom to accommodate future requirement changes.’‘

Marioni: High-end systems are: 

• where inconsistent ambient light cannot be controlled

   

• where parts cannot be consistently presented to the cameras
• where simultaneous multiple views, angles, fields of view of the same object are required
• where high throughput, intensive processing, and real time SPC data output are required
• where the customer requires custom user interfaces as well as images and data displayed in real time

and these are specific applications: 

• Tire Tread Identification/Differentiation: This is one station of a 6 camera / 6 station system where each cam