• Font Size:
  • A
  • A
  • A

Feature Articles

Smart Cameras: The Last Step in Machine Vision Evolution

by Winn Hardin, Contributing Editor - AIA

When companies like DVT first offered 'smart cameras,' acceptance was limited to niche markets such as distributed sensing systems for large area inspections - multi-stage production line monitoring, for instance. The systems were low cost, but most were written off as inadequate for all but the simplest vision applications. 

Today, with processing power and alternatives to RISC microprocessors increasing while component costs remain relatively stable, smart cameras are expanding to encompass a larger section of the vision market. 'We're seeing a migration of the functionality of what was once only PC based systems down to the smart camera level,' said Mark Sippel, product marketing specialist for vision at Omron (Schaumburg, IL). 'For example, some manufacturers are offering 2-D bar code reading and optical character verification on their smart cameras.'

As manufacturers jump on the smart camera bandwagon, the traditional paradigm of camera-with-processor is disappearing in an avalanche of new designs. Today, a smart camera seems to fit almost any system that doesn't use a PC, shoots for plug-and-play compatibility and comes relatively cheap. Starting at just over US $1000 and ranging upwards of US $6000, smart cameras are hitting the fabled price points while delivering the processing power that many experts have said will take machine vision out of the specialist field and into the commodity market.

Different approaches to 'smart'

The traditional 'smart' camera is a single box that houses a sensor and processing element. Communication lines were secondary and limited at best. However, as we'll see, many of the industry's leading manufacturers offer unique combinations of input, output, sensor and processor to provide unique smart camera solutions.

Sippel's Omron offers several smart sensor systems based on RISC microprocessors. Ranging from US $1200 for the low-end monochrome F10 to US $6800 for the color F400, the systems offer a variety of hardware configurations and algorithm approaches in an effort to meet a broadening scope of applications. For example, the F10 offers three fixed lens mounts, integrated lighting, remote camera head and amplifier (processing body) with digital I/O and RS 232/422 outputs. Its grey scale pattern search algorithms utilizing approximately 200 horizontal by 190 vertical pixels and narrow field of view (4.5 x 3.5 mm to 25 x 25 mm), make it appropriate for a absence/presence determination of small parts that will easily fit within the narrow field of views, while programming is done through push buttons on the face of the processor.

Omron F400
Omron's F400 represents one of the first color smart camera entries. Company officials say the system has opened up the plastics market to smart cameras where inspecting slight color variations is very important.

Stepping up to the F30 model, the system offers a more expensive wide-area lens (50 mm), but steps back to pixel counting algorithms for absence/presence and area measurements while allowing greater operator control through region of interest (ROI) definition through a laptop. Both the F10 and F30 models can store up to 8 'models' or representations of acceptable parts, making them effective at limited measurements on a diverse production line.
At the end of the scale, the F150-2 and F400 brings additional memory, DeviceNet capability (F150-2) and color (F400) to the table, allowing the operator to store 16 products with 16 regions of interest per image similar to general purpose machine vision system. 'The introduction of color into smart cameras at these price points opens up all new areas, such as the plastics industry, where color vision is imperative in determining between materials of similar hues. Before it had to be a large, expensive system. Now you can do it with these point color sensors,' Sippel said.

Enter the smart box

For example, SIGHTech's (San Jose, CA) Eyebot doesn't include a camera at all. 'Most smart cameras require you to use their proprietary camera technology. We've taken the intelligence out of the camera and put it in a little box about he size of a hard cover book,' said Francis Tapon, vice president of sales and marketing for SIGHTech. 'It just needs to be NTSC or PAL, and that gives us pretty much any video input. You could hook up an IR or x-ray camera to this system, and as long as it puts out an NTSC signal - which many of them do - then we can work with them.'

SIGHTech's Eyebot
 SIGHTech's Eyebot works with any NTSC- or POW-compatible camera including IR and x-ray.

According to Tapon, Eyebot's main discriminator is its ability to self learn without the use of complicated, memory-expensive graphic user interfaces (GUIs). The system is programmed either through the use of a laptop or via an external monitor and joystick device. An application specific integrated circuit (ASIC) using field programmable gate array (FPGA) technology gives the system enough processing power to meet many applications without significant user input. 'You basically connect to it and tell it to learn what's good. And it has to be most of the incarnations of good. In other words, if you're going to be inspecting bottles vibrating on the line, then you should teach it bottles while they're vibrating down the line at high speed -- in other words, all variations,' Tapon said.

Eyebot offers three outputs: an integer between 0 and 99 that scores the product and two optically isolated outputs for PLCs. 'Most manufacturers will say we need lots of outputs and data collection. But in the end, what people really want is just a go, no go.' Tapon added the system has received significant interest, claiming clients such as IBM, Raychem and Raytheon.

Let hardware do the hard stuff

In this quest to expand the end user market through simpler integrated inspection solutions, Pulnix America (Sunnyvale, CA) will soon enter a new contender - the ZiCAM. According to product marketing manager Manish Shelat at Pulnix, the ZiCAMs strength lies in its flexibility, strong processing capabilities, simple operation and compact size. 

Pulnix America's ZiCAM makes use of a hardware neural network to make learning and inspection as easy as possible for the operator.

ZiCAM stands for Zero instruction CAMera and is based on the IBM/Silicon Recognition ZISC chip (zero instruction set computer). 'We call it zero-instruction set because it's a hardware-based neural network camera,' explained Shelat. During a walkthrough, a 640 x 480 pixel CCD based on Pulnix's 6700 series CCD camera captures the image and feeds it into an internal proprietary subsystem called MUREN (MUltimedia Recognition ENgine). MUREN conducts several feature extraction functions, including histogram, x profile, y profile, x and y profile and pixel subsampling. From there, the data is converted into a 64-bit vector and fed into two ZISC processors, each holding 36 separate computational units called neurons -- hence the neural network.

ZiCAM with four ZISC processors (144 neurons) will be available as an option. ZiCAM also has a built-in 170MB microdrive (340MB optional) that can save 550 VGA resolution images. No other Smart Camera in the market has this feature.

Through training the system on real images, each neuron establishes numerical relationships between each computational element. The 'sum' of these values defines the part, allowing for rugged automated inspection. Training is completed using a laptop that is disconnected during operation. Training the ZiCAM is accomplished by stepping through saved images and classifying these images as good, bad, and other predefined categories. An end-user builds the ZiCAM's knowledge set via this simple training process.

The complex hardware is hidden behind a simple good/bad GUI, while action and reporting are handled through a pair of RS 232 outputs and a 15 pin digital I/O connector for PLC communication, etc. Monitor and PS/2 connectors are also included in the 71.5-mm x 71.5-mm x 155.4-mm camera. The ZiCAM will also be available in a remote head version for machine vision applications where camera mounting space is limited.

'ZiCAM is excellent for parts sorting, presence and absence detection, parts and features identification, and label and character verification,' Shelat said. Currently, ZiCAM is undergoing Beta testing. Its expected to hit a machine vision store near you by early fourth quarter at a list price of US $5800.

Let software rule the day

Taking the opposite approach, smart camera pioneer DVT Corp. (Norcross, GA) believes that a single hardware platform with upgradeable software library offers the best approach to smart camera technology. 'We need to be at a photoelectric price with a vision system. And while [the photoelectric sensor] can be dumb, we need to be a very smart product; something that doesn't need a light and can gather whatever information is necessary. We look at ourselves as not a vision system, but a data source,' said DVT CEO Robert Steinke. 

DVT's latest hardware platform, the series 600, builds on the speed of Motorola's PowerPC embedded processor for a 10X speed improvement over the previous 700 series. This latest iteration offers a standard 640 x 480-pixel CCD with 12 configurable I/O, and ethernet connectivity in addition to RS 232/422 connectors. DVT's SmartImage sensors  decouple the path from the CCD to memory to allow processing on a portion of the overall frame while another image is read from the CCD. 

DVT's smart camera approach is to keep the hardware platform up to date through the use of a powerful processor (Motorola's PowerPC), and continually update the transparent software.

According to Mike Schreiber, director of applied engineering at DVT, the camera is trained through a 'windows' GUI for user familiarity. However, by keeping all processing within the camera, DVT engineers retain control that can sometimes be lost when adapting machine vision algorithms to the Windows operating system. DVT also includes 1D and 2D barcode readers and OCR algorithms on it's SmartImage sensor. 

With this kind of functionality, Schreiber said its easier to say what smart cameras are 'not appropriate for and that's nothing. They're in every industry. We're seeing [smart cameras] open up the end user market on the plant floor. In the past, the market was semiconductors. Machine vision relied on 70 percent semiconductor and 30 percent automotive. Now it's across the board in many industries.'

Schreiber expects that DVT's release of SmartLink in the near future will increase the ubiquitous nature of smart cameras. With it's own PowerPC chip, the SmartLink box will allow the monitoring of  multiple (up to 16) series 600 smart cameras over Ethernet without the use of a PC. Since each camera is connected via an ethernet with its TCP/IP stack and IP address, the cameras will be accessible from remote locations across intranets or even the World Wide Web, if IT wants it that way.

Compromise with power

While DVT's software is downloadable free from the company's Web page, Cognex is loading it's In-Sight 2000 smart camera entry with the 'geometric pattern matching' techniques first debuted in PatMAX. 'It's a classic vision sensor,' explains Cognex's senior manager, end user marketing, George Blackwell. 'We looked at the market and saw that there were vision systems with high end tools and a lot of performance that needed integrators. At the other end, we saw smart cameras and some low end vision systems. Typically, these were fairly easy to use, had a nice interface and very low cost, but didn't use sophisticated tools. Sometimes they used binary processing or grey scale for simple pattern matching. Our take is that even if you're doing a simple application, there are complicating factors, such as vibration, lighting changes, part surface finish changes and so on. There were always something like the last 10 or 20 percent that was difficult to achieve with a smart sensor.'

With the problem so defined, Cognex decided on a standard 640 x 480 CCD using DSP processors to keep power consumption and heat dissipation to a minimum. The result is a box without any moving parts that can stand up to rugged plant environments, and give the piece of mind of an all electronic system, Blackwell said. Over that, Cognex has added its latest spreadsheet programming environment, controlled by a handheld device similar to a game pad  for less than US $5000.

Cognex's In-Sight2000 uses a spreadsheet programming environment to bring standard machine vision functionality to the smart camera realm.

A new, distributed vision

The appeal of smart cameras can hardly be denied. According to Chris Boeri, an electrical engineer with RJR Polymers (Oakland, CA), a PC-host system is too much for some applications. 'We wanted to make integration of the vision system as simple as possible, and the In-Sight product seemed to fit the application nicely,' explains Boeri. 'Since we only needed feedback on the X, Y, and theta, we only needed a single camera system. In-Sight made it nice and easy to pass data over the RS-232, and since it's a standalone unit, it saves us from having to integrate and maintain a separate PC. A PC-based vision system would have been overkill.' Cognex claims to have already booked more than US $1 million in sales on from the In-Sight system since March 2000.

According to Omron's Sippel, smart cameras are an idea who's time has come. 'One thing that has had to happen is a shift in philosophy. Instead of trying to do all inspections at one time and in  one location, and trying to maximize as much as you can in one system, people are breaking the inspections out across an entire process. When you drill a hole, you inspect a hole, instead of waiting to the end of the line and inspecting during final assembly. Once you accept that thought processs, everything becomes more affordable. Not only to catch bad parts before you added a lot of value to them, but you save yourself on broken tooling as well, discovering a broken drill bit as soon as the first part comes through without the hole.'

About the Author: Prior to accepting his position as managing editor of Wireless Design Online, Winn Hardin spent several years covering the vision industry for a variety of leading trade publications.






There are currently no comments for this article.

Leave a Comment:

All fields are required, but only your name and comment will be visible (email addresses are kept confidential). Comments are moderated and will not appear immediately. Please no link dropping, no keywords or domains as names; do not spam, and please do not advertise.

First Name: *
Last Name: *
Your Email: *
Your Comment:
Please check the box below and respond as instructed.

Search AIA:

Browse by Products:

Browse by Company Type: