When does it make sense to use color machine vision instead of or in addition to monochrome? And how do cameras “see” color anyway? Experts from Dalsa Corp., Waterloo, Ontario, answer these questions and provide insights into the origins, evolution, and future of color in machine vision.

Can a machine see in color like a human?

A machine calibrated to the average human response to color (to a “standard observer”) can produce consistent responses to color variations in a controlled setting. This “calibrated color vision” is useful for measuring and matching colorants in paint, plastics, fabrics, and other materials. We don't think of it as “seeing” color like a human, except perhaps as a philosophical exercise.

What are the basic differences between machine and human vision when it comes to color?

Human color vision is extremely versatile in that it can reliably extract information about objects despite huge variations in illumination and view. For example, humans can reliably judge fruit based on color, distinguishing ripe from unripe or bad fruit despite varying illumination and viewing perspective. Human color vision has inherent mechanisms that “factor out” variations in illumination and view, things we don't know how — or don't bother — to put into machine color vision. Human color vision, on the other hand, is relative in that nearby colors influence the perception of a particular color. It also has low resolution (a fact used to transmit color in television with very little bandwidth) and differs widely between individuals, making it not a very good measuring tool. Machine color vision, by contrast, is not influenced by nearby colors, can have high resolution, does not vary much from machine to machine, and thus is a good measuring tool.

How do you see color machine vision evolving in the future?

Certainly the use of color will expand in machine vision. Color provides much more visual detail than monochrome grayscale and adds a new dimension in analyzing real-world data. For example, color is increasingly being adopted in bank note inspection applications for scanning and processing. In some Asian countries, the governments require color because people use personal seals rather than signatures in issuing personal checks. The seals usually employ red ink, which has poor contrast in a monochrome system. In such cases, it's typical for banks to require a color report to confirm authenticity.

Another example is circuit board inspection in which color cameras are used to identify oxidized copper wires that are otherwise difficult to see in a monochrome system. Better color fidelity, lower cost, and ease of use are the primary market drivers and we foresee new technologies being developed in the near future to address these needs.

When does it make sense to use color over monochrome?

Aside from the obvious applications where the color of an object needs to be evaluated in some way, color can sometimes help to make an inspection situation easier by facilitating the identification of objects, such as in the verification of fuse values in car fuse boxes.

Should color eventually be used in all machine vision applications?

No, because there are some things that monochrome cameras will always do better than color cameras. Take resolution and speed for example. You will always find more choice in high resolution and high speed if you are shopping for a monochrome camera. In addition, there are many cases where color images do not offer any advantage over monochrome images in resolving a machine vision problem. For the typical inspection application where it's necessary to detect defects such as cracks or scratches — in other words, discerning variations in brightness on an object's surface — monochrome usually outperforms color.

How does color detection work in terms of hardware (cameras)?

In a charge-coupled device color (3CCD) camera, color is selected using a prism-based interference filter that splits the incoming light into its primary (red, green, and blue) components. Each of the three colors is then detected by a dedicated CCD and the final color image is reconstructed by combining the outputs from the three detectors. All three color images are captured at the same object spot and at the same time. In a tri-linear color camera, three linear arrays are fabricated on one single die and coated with corresponding color filters. These are absorbing filters that use dye or pigment. In the tri-linear camera, the three linear arrays detect a slightly different field of view, necessitating spatial correction in the reconstruction of the color image.

What role does software play in color detection?

Most color machine vision systems use a mix of hardware and software to detect colors. For point or “spot” color measures, a hardware-centric solution is fine. More sophisticated detection systems, on the other hand, tend to use more software, giving designers and users greater flexibility. In most systems, the power of the software lies in color “classifiers,” decision-making modules that detect colors and then assign color pixels to a class such as “good” or “defect.” A good classifier has some tolerance to illumination changes, is quick to train and run, and reliably assigns pixels to their correct classes. Classifiers are an area of continuing development and differentiation among competing solutions.

What are the major challenges regarding color detection?

Machine color vision systems do not have “color constancy,” as humans do, so the lighting must be specified and controlled to provide robust color detection. A patch of known color in the camera's field of view can be used to provide some constancy, but the ability to compensate for illumination changes is still limited. If color distance measures are to be made, for example the diameter of a yellow object on a blue conveyer belt, then a suitable choice is to use either a three-sensor camera or a single-sensor Bayer pattern camera with reduced resolution and careful pattern decoding.

What types of color machine vision cameras exist, and what are their pluses and minuses?

The main types of color cameras used in machine vision applications are 3CCD, tri-linear, and Bayer pattern cameras. 3CCD has excellent color registration and can be applied to the majority of applications. However, the cost is high due to the design. Tri-linear provides high performance and has advantages in terms of its low cost. It can be used in many applications such as flat surface inspection. However, spatial correction cannot be done properly in certain applications that involve rotating or randomly moving objects. Bayer pattern cameras round out the mix and offer the lowest cost solution. These cameras tend to be used in lower-end applications as they suffer from reduced color precision when compared to 3CCD and tri-linear cameras.

Special thanks to Dalsa experts Robert Howison, project leader, OEM applications group, Ben Dawson, director of strategic development, and Xing-Fe He, product manager, for this month's tips. For more information, visit www.dalsa.com or call (519) 886-6000.