CMOS Sensors Offer Vision Designers New Capabilities

Dalsa Corporation offers this comparison of performance characteristics between the two main imager types used in industrial vision applications: charge-coupled device (CCD) and complimentary metal oxide semiconductor (CMOS) imagers.

Unfortunately, efforts to optimize logic-based CMOS manufacturing for imaging sensors has proven more difficult than expected. CMOS sensors suffer from non-uniformity because each pixel has its own charge-to-voltage conversion, while a CCD sensor uses a single charge-to-voltage conversion for every pixel. This means that if a charge to voltage conversion circuit doesn't work on a CCD, then the whole sensor is bad and is scrapped, while a CMOS sensor may have several converters that are bad, but the sensor is still used because the rest of the pixels perform as expected. This challenge, however, allows the CMOS sensor to be selectively read, also referred to as windowing, meaning that the user can configure the sensor to only read out certain rows and/or columns of pixels or even individual pixels, increasing the frame rate of the sensor by effectively reducing its size. When camera designs containing programmable ‘‘full descriptors’‘ are added to a CMOS sensor, the size of these windows, along with many other performance characteristics such as gain, exposure time, etc., can be changed from frame-to-frame, giving the user maximum control of the sensor and leading to small-form-factor, highly-configurable, high-speed CMOS camera designs.

‘‘We used to use CCDs in the past, but then we switched to PixeLINK's [CMOS] camera because of the FireWire interface and the CMOS imager capabilities,’‘ explained Pierre Aubrey, executive vice president of ShapeGrabber Incorporated (Ottawa, Ontario, Canada), supplier of the 3D imaging/scanning systems that use a laser and CMOS sensor to create 3D maps of objects. ‘‘In addition to a high-speed consumer-based interface, CMOS gives us a fair bit of control over things like windowing, and other parameters...exposure control, that kind of stuff. We also benefit from a price point of view.’‘

According to Joel Bisson, president and CEO of CMOS camera supplier PixeLINK (Ottawa, Ontario, Canada), a megapixel CMOS machine vision camera can cost $1800, while a similar CCD camera would cost $3000.

Configurable imaging systems
By combining CMOS sensors with gigabytes of local memory, Integrated Design Tools Inc. (Tallahassee, FL) fully realizes the speed of CMOS sensors for applications such as particle image velocimetry (PIV), fluid flow analysis, high-speed impact analysis and high-speed inspection. ‘‘As we make the region of interests smaller, the frame rate goes up accordingly,’‘ explained Luiz Lourenco, IDT's CEO. ‘‘By adding a processor and configurable memory up to several gigabytes inside the camera, the system can take advantage of the full frame electronic shutter capability of the CMOS sensors to acquire back-to-back frames in less than a microsecond.’‘

Like PixeLINK, IDT's cameras use a high-speed interface (USB II), but the data typically is stored on the camera for later download to maximize the speed of the camera by eliminating bottlenecks between the camera and image processor. The next generation camera will offer gigabit Ethernet for even faster downloads, and real-time downloading. By combining the X-STREAM cameras with onboard microprocessor, a PC and external timing hub, which acts mainly as an enhanced pulse generator, the high-speed imaging system can synchronize multiple cameras and change the camera's gain, exposure and frame rate based on external triggers that can include features extracted from real-time images. ‘‘For instance, you can do 1000 fps before an impact, and 5000 fps after the initial impact.’‘ The X-STREAM line is designed to work with National Instrument's (Austin, TX) LabView for general image analysis and Math Lab for customized algorithm development.

CMOS sensor performance, pitfalls
PixeLINK's Bisson recommends that end users be familiar with how the camera manufacturer deals with CMOS three major technical challenges: fixed-pattern noise (FPN) and photoresponse non-uniformity(PRNU), and parasitic sensitivity. CMOS camera suppliers use flat field correction that adjusts the gain and offset control for each pixel to counter FPN and PRNU, both of which should be addressed by the manufacturer at the factory. Sometimes, pixels cannot be corrected and as such are identified as defective.  Manufacturers typically fix this by replacing the pixel with values from a neighboring pixel, which is less effective, or using an average or gradient based on neighboring pixels. The more pixels values taken into account, the better the final image, but the more space required on a nearby field programmable gate array (FPGA) to perform the correction. Bisson adds that this is less of an issue with larger arrays because a single pixel is not as likely to affect the image quality as the array increases. ‘‘Some manufactures will publish a map of the defective pixels per sensor,’‘ Bisson said, adding that the number of defective pixels is usually less than 0.01 percent of the total array. Manufactures should also provide a software tool for on-site flat-field correction that will take into account uneven lighting at the application and correct accordingly.

Parasitic sensitivity occurs when a pixel continues to collect charge after shuttering. This is more of a problem with systems installed in locations with excessive ambient light or with systems that do not strobe the light source. CMOS sensors typically use one of two shutter methods: rolling shutter or global shutters. Rolling shutters are faster, reading out one row of pixels while exposing another, but are better suited for step and repe