Time to revisit machine vision?

Figure 1 Multiple cameras can eliminate the need for precision x-y motion stages in optical inspection systems. Courtesy of CyberOptics.

Knowledge-based system advances are piggybacking on innovative hardware that AOI equipment makers have adopted—including multiple-camera configurations (Figure 1), high-resolution charge-coupled devices (CCDs), and advanced illumination schemes—to build capable systems able to identify a variety of PCB fabrication problems (Figure 2). But one major obstacle to widespread adoption remains: "Customers really don't want to have a dedicated engineer working on the system all the time, because it drives up the cost of operation," says Jim Fishburn, general manager of the Inspection Dept. at Omron Electronics (Schaumburg, IL; www.omron.com). Nevertheless, AOI systems still require manufacturers to keep engineers on the payroll to develop and maintain working test programs.

Early vision scientists manually programmed ideal reference image templates and then used gray-scale correlation and pixel counting to compare the templates with real-world images. Many factors complicated this conceptually simple process: lighting, background, color, reflections and shadows from surrounding surfaces, and acceptable component-size and process variations from one board to the next.

For AOI users, failing to predict image variations correctly results in a high false-failure rate, or at least imposes time-consuming manual intervention for AOI test-program generation and maintenance. Manufacturers have addressed these issues by adding more-rigorous algorithms, adaptive models, and easier-to-use operator interfaces (Figure 3) to their products, moving optical inspection systems away from the "golden board" approach toward a more "intelligent" approach that offers built-in process knowledge.

The move essentially decreases the time required to generate and maintain AOI test programs from days or even weeks to a few hours or less. The bottom line is that AOI systems are getting easier and less costly to use for improving first pass in-circuit test (ICT) yields.

Intelligent AOI systems are beginning to reduce the engineering time necessary to develop an effective image-analysis program. All AOI programs share the same goal: catch defective PCBs and pass good ones, without rejecting good boards or passing bad ones. Most systems developers work through three phases to develop a new test program: program development, debug, and tuning. The tuning phase consumes most of the programmer's skill and time.

Programming projects commonly begin with CAD interpolation. Many AOI systems use vision libraries to store process knowledge, test procedures, normal variation data, component models, and other information used during image processing. AOI-specific drivers translate information such as component type, size, and position data from the CAD system.

Often, a user needs only a mouse click to import the data into the AOI system or an off-line programming station. In some instances, however, differences in customer terminology and vendor naming conventions may require users to intervene manually. Most systems flag unknown or unidentified parts or descriptions and then prompt users to identify component type and make the assignment to an appropriate library component.

Using automated software development tools, many AOI vendors have reduced the first stage of programming to as little as 10 or 15 minutes. Depending on board complexity, test procedures to be employed, and the number of manual interventions required, users may spend another 10 or 15 minutes mapping component assignments to the library.

The second phase of developing a program, debugging, may also take as little as 10 to 15 minutes. Operators typically show the system a single good production board to validate that the correct revision of the CAD file was downloaded and that registration marks are correctly located.

Next, the program must be fine-tuned, which can take an expert anywhere from a few hours to several days, depending on the complexity of the PCB. Tuning validates that the program works over the first hundred or thousand production boards. To develop a working program that is resilient to variations and normal production conditions, the programmer must refine the pass/fail criteria to accommodate real-world variations.

"The real evil of the tuning phase is that it could take forever," says John Arena, marketing manager for inspection and test at Teradyne (Waltham, MA; www.teradyne.com). "Every time the system encounters a variation it hasn't seen before, the user has to judge it as good or bad and tweak the program. So, in many instances, it's an ongoing process."

Thanks to higher-power processors, suppliers can make better use of knowledge gained from their inspection experience. Teradyne, for example, will introduce its Optical Process Test, a complement to AOI, during the 2003 APEX show (Anaheim, CA, March 31-April 2; www.goapex.org). "OPT ships with built-in process knowledge to determine if a board is good or bad," Arena explains. "This significantly reduces the amount of time users have to spend fine- tuning the system. Users just download the CAD data, wait a couple of hours, and then fine-tune the system by showing it 10 to 20 production boards."

Other AOI suppliers have built process knowledge into AOI systems using different approaches. For example, ViTechnology (Haverhill, MA; www.vitechnology.com) uses geometric pattern recognition from Cognex (Natick, MA; www.cognex.com) in its Vectoral Imaging systems. The company combines component model libraries with a CAD definition of the board to originate programs in terms of position, component type, and board layout.

The libraries use component renderings instead of actual images, extracting only the important features on each component such as length, width, and the number and position of any leads. "These rendered models are more transportable between various applications because they exclude the component details that are likely to vary," explains technical sales manager Jean-Marc Peallat.

Figure 3 Easier-to-use interfaces are driving an evolution from idealized systems to knowledge-based systems and decreasing the time and skill required to generate and maintain AOI test programs. Courtesy of Cognex.

Agilent Technologies (Dublin, Ireland; www.agilent.com) also uses feature extraction to pass certain characteristics of the devices under test through its AOI system. Additionally, the company's system employs self-learning algorithms that mix neural network with multiple variable classifier technology. "We don't store images in our library," explains Thorsten Niermeyer, senior product manager. "We store mathematical representations of each feature and use multiple comparisons to inspect the unit under test."

Niermeyer explains that while the initial fine-tuning of Agilent's AOI system may require showing the system anywhere from 10 to 100 boards, tuning gets faster as the library evolves. For example, if a new board design uses the same components as an earlier design, the system re-uses the data in the library. "You don't have to start from scratch," Niermeyer explains. "All the fine- tuning that you have done before will be re-applied to new programs."

Photon Dynamics (San Jose, CA; www.photondynamics.com) takes a different approach. It prefers to start from scratch developing models. Rather than tweaking a library full of canned models to handle real-world variations, the company's system uses interactive artificial intelligence developed at Intelligent Reasoning Systems (Austin, TX), which Photon Dynamics acquired in 2001. "We train on real boards—learning what's acceptable and what's not by experience," says corporate VP of research Mark DeYong.

The training process involves exposing the system to multiple examples of good assemblies, and possibly bad assemblies, to build a knowledge-based model. The advantage, according to DeYong, is that the system can actually find defects that have never been specified by the operator. For example, he explains, "there's no way to arbitrarily define a crack. Yet, an adaptive system can detect this and notify the operator to establish a new class of fault."

CyberOptics (Minneapolis, MN; www.cyberoptics.com) manufactures self-learning AOI systems that use statistical appearance modeling to simplify training, reduce the amount of time required to deal with false calls, and make the machine more usable for line operators. While the system's software aids in variation prediction, it still requires programmer intervention. Yet, the more boards the system is exposed to, the better its ability to discriminate. "After a sufficient number of examples, the system runs on its own," says technical marketing manager Andy Yates. "However, it still requires periodic operator and programmer intervention."

One concern with self-learning AOI systems, especially in very low-volume/high- mix environments, is that the run could be over by the time the system is fully trained. But Yates points out that "operator interface improvements have made knowledge-based systems as quick, if not quicker, to program than idealized systems."