GigE Vision expands in machine vision

The GigE Vision interface standard brings several benefits to machine-vision inspection of PCBs (printed-circuit boards), semiconductors, and wafers. A flurry of product introductions this year has demonstrated the growing popularity of the interface in machine-vision cameras, from small embedded boards to two-box, 12-Mpixel cameras (see Machine-vision cameras get GigE).

In this example of a manufacturing line, images are sent over an IP network using standard Category 5 or 6 Ethernet cables to a control room where rack-mounted servers process them in real time. Courtesy of Pleora Technologies.

These aspects of GigE also make it possible for manufacturers to set up control centers where images are processed in multiple, clustered PCs, said Eric Carey, R&D director for Dalsa Montreal. “Some of our customers are considering using a GigE Vision link to transfer images to a control center for processing,” he said.

But when you send GigE camera data to an Ethernet switch and then to the PC, bandwidth can’t be guaranteed and packet loss is likely. “With higher-bandwidth cameras, you need to be careful about aggregating camera bandwidth so you don’t overload the link,” Carey said. “So, most customers are presently implementing a point-to-point connection.”

Ethernet’s multicast capability is what makes it possible to decouple image acquisition at the camera from image processing at the PC, said George Chamberlain, president of Pleora, a co-founder of the GigE Vision standard. “In a LAN, multicast capability lets you send video data simultaneously from one source to multiple locations,” he said. “Processing, archiving, monitoring, or even additional processing, perhaps using multiple PCs, can be done at multiple locations.” He added that using FPGAs (field-programmable gate arrays) instead of DSPs (digital signal processors) to perform image processing can make a multicast solution more scalable, less expensive, and simpler to develop.

GigE Vision is also scalable to accommodate networks operating on the 10 GigE standard, and it eliminates frame grabbers, which can cost as much as $3000 depending on functionality, said Ravi Guntupalli, business manager for Princeton Instruments’ imaging division. He explained that upgrading a network to handle GigE Vision in the first place and coping with the loss of frame grabbers can each pose separate design challenges.

Since frame grabbers traditionally handle image processing, eliminating them increases the load on the host CPU, adds image-processing CPU cycles, and decreases the number of CPU cycles available for object recognition and registration. Guntupalli said that using cameras with built-in image-processing capabilities can help reduce the CPU load considerably. In Princeton’s MegaPlus line, for example, those capabilities include RGB color interpolation, flat-field formalization, and defect correction. Implementing them with FPGAs instead of DSPs, as Princeton has, can reduce time, cost, and programming effort, he said.

Alternatively, the task of converting GigE Vision packets into usable images can be offloaded from the host CPU by using a general-purpose coprocessing board and a NIC (network interface card) on the PC’s PCI Express bus, said Carey. “Many of our customers are still using typical analog cameras with a frame grabber and a dedicated link, so aggregating is a new concept,” he said. “It becomes easy to scale a system and easy to try to do too much, so you need to be aware of the average bandwidth that can be sustained on the link.” In many inspection systems, this bandwidth is still only about 10 to 15 Mbps, so multiple cameras can be handled with a NIC.

Of course, modifying an existing network to take advantage of the GigE Vision standard can be an issue with cameras, because sensors are more expensive components than interface electronics, yet they don’t need to be swapped out nearly as often, said Guntupalli. One way to address this problem is to place the bus-interface electronics in one box and the more-expensive image sensor electronics in another. You can swap out the interface and use the camera head with whatever interface you choose. “At resolutions of around 11 or 16 Mpixels, sensor prices will not be coming down that much because volumes are lower, and because the amount of silicon used is a lot more,” Guntupalli said. Princeton is one vendor that has taken this approach in its MegaPlus line.

Although many manufacturers have accepted the GigE Vision standard and are encouraging widescale adoption, adoption rates have been slow to match initial expectations—as is often the case with new standards. Princeton, for example, expects customers who have complex inspection setups on their PCB production lines to stay with Camera Link because of its high data throughput, said Guntupalli. “But in offline semiconductor wafer inspection, customers have more incentive to use GigE, since their throughput needs are lower and they can benefit from its reduced cost, longer cable length, and ease of use,” he said.