Cognex Corp., Natick, Mass.

With so many vision systems available today, identifying the most suitable for a particular application can be daunting. Vision parts that just perform necessary tasks are not enough; there are several factors that must be considered for successful deployment.

First consideration: Setup

Vision applications don't usually require elaborate runtime interfaces, but operators often interact with vision systems during part change-overs, to alter tolerance parameters, and to determine causes of failure.

Better vision systems allow quick and easy configuration of these and other application facets without coding in Visual Basic or proprietary script-based language. Some vision software also includes network management tools to simplify remote administration of multiple systems, including backup, image playback, firmware upgrades, and context-sensitive help.

A couple tips: Operator interfaces that display images allow immediate analysis of failed parts, as well as pass/fail statistics, to help operators quickly identify trends. Some vision machinery can also be modified and turned off and on by operators.

Part location tools

Machine vision requires software to find parts within the camera's field of view. Setup with this software is typically the first step in any vision application, from simple robotic pick-and-place operations to assembly verification tasks. It's also the most critical step, as it determines application success or failure.

Locating parts in an actual production environment can be challenging. First, vision systems are trained to recognize parts based on a model image. However, even tightly controlled manufacturing processes vary in the way parts appear to the vision cameras. Therefore, vision part-location software must be intelligent enough to compare model images to actual objects moving down a production line, regardless of which side of the part faces the camera, its distance from the camera, shadows, reflections, line speed, and normal appearance variations.

Image preprocessing

Preprocessing tools are software that alter raw images to emphasize target features and minimize unwanted ones. This prepares images for more powerful vision tools and can significantly improve overall robustness. Preprocessing tools can increase the contrast between the part and its background, mask insignificant and potentially confusing image features, eliminate hot spots reflecting off of surfaces, and differentiate smooth and rough textures.

As we'll now explore, image-preprocessing tools also optimize trained models by sharpening the edge contrast of characters and filtering out extraneous background in the image — so markings on products are read more reliably.

Character reading and verification

Whether vision parts are reading stamped alphanumeric codes on automotive parts or verifying date and lot code information on medicine bottles or packages, several capabilities are paramount for character reading and verification.

Statistical font training

This capability builds a font by learning models of characters that appear in a series of images. The images should include multiple instances of each character, and span the full range of quality likely to occur in production. The resulting font is tolerant of normal variations in print quality, whether due to poor contrast, variable locations, degradation, or stroke-width variations. Unless a designer knows in advance that every code will be marked with the same quality seen in the reference images used to teach character models, statistical font training can be crucial to the success of reading or verification applications.

Instant image recall

This capability enables line operators and technicians to quickly and easily view failed images on a display. Whether a camera jarred out of position or a damaged label causes failure, it is important to know immediately why failure occurs, so corrective action can be taken.

Consider a packaging plant, in which container materials, labeling equipment, printing methods, and ambient lighting can vary considerably over time. Here and in similar applications, designers should perform tests on large samples of good, marginal, and poor-quality labels to see how the vision performs under variable real-world conditions. Because character positions can shift from label to label, it's also a good idea to enlarge the region of interest around the character string. This will help determine how reliably the vision system's reading and verification tools operate within a larger search region.

Next Page: Repeatability and codes