Machine vision looks good

Ben Dawson, Dalsa Coreco -- Test & Measurement World, 11/1/2005

Machine-vision technology has been used for more than 30 years, yet is perceived as being difficult to set up. This perception comes from vision products that were difficult to use, often underpowered, and required a vision expert. Fortunately, a new generation of vision products are changing this perception and bringing wider acceptance for vision technology.

Figure 1 shows an assembly process for electric drills. A conveyor moves printed-circuit boards from manufacturing to the machine-vision system, which detects the board, checks that each board is the correct type, checks for missing components, and reports the board's position and orientation to a robot arm controller for correct insertion into the drill assembly. A few years ago, this task would have required an expensive vision system and extensive development.

Most machine-vision systems have a way to position the parts in the camera's field of view (the conveyer belt here), a "part-in-place" sensor to detect that a part is present, lighting, a camera, a vision computer to do the inspection, and some outputs based on the results of the inspection.

The part-in-place sensor has developed from the lowly photocell to be smarter and easier to use. Sensors are now available for specific types of materials, such as reflective or matte surfaces. Some "smart sensors" include simple processing that, for example, allows them to detect and sort different colored parts.

In the past, vision systems used cameras designed for surveillance or television. These cameras were unsuitable for making fast and accurate measurements of part dimensions. Machine vision now uses specially designed cameras that are fast and provide the high-quality images needed for inspection and measurement.

Lighting is a key component in a machine-vision system—you can't inspect or measure something you can't see. Carefully designed lighting enhances the features in a part that you want to inspect or measure while suppressing visual features that interfere with the inspection or measurement. For example, to find and count small particles on a surface, you can use a ring of lights pointing nearly parallel to the surface. This "dark field" illumination makes the surface black while particles catch the light and appear bright white.

The advent of intense, inexpensive, uniform, and reliable LED lighting has made setting up a machine-vision system much easier. You can find hundreds of lighting products specifically designed for machine vision. Most lighting vendors will recommend appropriate lighting for your machine-vision task, so you don't have to be an expert.

The vision computer can be a specialized processor inside the camera—a "smart camera"—or a separate, specialized computer, typically based on personal computer technology (Figure 2). The advantages of this approach include increased computational power and a huge base of available software and experienced programmers.

For advanced applications that call for data rates of tens of megabytes per second, you can turn to faster, more powerful dedicated vision-processing boards. Although such boards are not new, today's versions use DSPs and FPGAs for extraordinary computational power.

Most vision computers—from smart cameras through high-end processor boards—include some form of digital I/O. Lower-end computers generally provide one or two input lines to trigger image acquisition and two or three output lines to drive part sorting. More capable hardware features more I/O lines and additional I/O protocols, such as Ethernet, FireWire, USB, and serial and parallel ports.

Some newer vision computers integrate what amounts to a small programmable logic controller (PLC) into the vision computer. For example, our vision computers provide separate logic that delays the part-sorting output signals.

An increasing number of standardized machine-vision components are making it easier for vision designers to assemble a machine-vision system. Components for part positioning are standardized already, and lighting, optics, and sensor (camera) components are fast moving in that direction. Standard vision components reduce deployment efforts—you get a familiar mechanical or electrical interface.

Vision application engineers understand an inspection system's lighting, optics, algorithms, and computation capability, but they don't necessarily understand the unique characteristics of a particular task. System development often bogged down while the vision application engineer educated you about the system's capabilities and you educated him or her about what you wanted to accomplish. Standard-component specifications embody much of this specialized knowledge and so facilitate this communication.

In addition to the many hardware advances, sophisticated machine-vision algorithms are also improving vision-system performance. For example, most machine-vision programs start by locating the part within the image, which they accomplish using a search algorithm. Machine-vision vendors are competing to improve the ease-of-use, speed, accuracy, and robustness of their search algorithms. Figure 3 shows a search algorithm setup screen from Dalsa Coreco's Sapera Processing software. Speed and accuracy are easily quantifiable. Robustness is somewhat more elusive.

One measure of "robustness" is how well a vision system ignores acceptable variations. Manufactured parts exhibit such variation, and "naturally manufactured" parts, such as apples or oranges, have even more. Robustness also indicates how well the system tolerates "distractions," such as changes in lighting conditions, unexpected objects in the field of view (an operator's hand, for instance), or changes in a part's reflectance. These conditions depend on the individual situation.

The more robust an algorithm is, the easier it is to use. If, for example, the algorithm can tolerate or compensate for some changes in lighting, then less effort is needed to shield the vision system from stray light and to control variations in system lighting.

Perhaps the most important factor in ease-of-setup is how the hardware capabilities, such as inputs and outputs, and the processing algorithms are presented to the user. Algorithms such as search or convolution (for spatial filtering, pattern matching, edge detecting, and so forth) have always been encapsulated into subroutine libraries. A well-designed subroutine library offers high execution speed, compact memory, and consistent execution time—important in real-time applications that press the limits of hardware performance. A major disadvantage is the modest-to-long learning curve required to use a library, and a subroutine library can contain only a limited amount of prior knowledge about a particular machine-vision application.

A key advance in ease-of-setup was the development of graphical programming packages. These packages present the vision system's software components in a graphical user interface so that you can design your application by connecting components rather than by programming. Not having to write code and having help online greatly reduces the system development time. For example, library-based applications typically take months to set up, while ones based on Dalsa Coreco's Wit or Sherlock packages, for example, can be done in weeks. There are still high-performance vision applications, such as inspecting large LCD panels, that require the performance of a subroutine library, but most applications can be implemented more quickly with a graphical programming package.

The latest and, in my opinion, hottest development in vision software is to tailor it for specific classes of applications. For example, our iNspect vision appliance is designed for inspection requiring single or multiple views of a part, while the iLabel vision appliance inspects labels on boxes, bottles, cans and similar items. Limiting the problem domain allows the embedding of a large amount of application-specific knowledge into each product.

Embedding functionality in this way can permit a much simpler user interface to the algorithms and hardware resources. Each specialized product's interface consists of panels, somewhat similar to a Windows wizard, that walk you through the problem (Figure 4).

By precisely matching system and application, you can often set up and verify a vision application in hours and with minimum help, although you may still need help with lighting and lens selection. These specialized products are designed so that users who are not vision experts can quickly build test and measurement setups.

Machine vision is becoming an increasingly important component in test strategies. The technology is being driven by processor performance levels, the challenge to improve algorithms, and the creation of standardized products. System development has recently started to emphasize ease-of-setup, as this is a critical factor in using machine vision for efficient and effective manufacturing.