Model Management in the System Generating Vision Inspections

accordingly. However, they are not able to build The article presents a possible approach to model management in a knowledge-based system with distributed knowledge. The knowledge based system is designed to automatically generate vision inspections. The proposed model management mechanism extends system reasoning power without increasing system complexity. The model management knowledge is structured similarly to knowledge for inspection synthesis. Consequently, important advantages of knowledge organization like sharing, flexibility, independence, etc. are preserved for both kinds of knowledge. and manipulate models if these are not provided externally (see [6, 51). The CAD models also do not contain some important information, for instance, explicit description of object defects as they occur in training images. Presented model management overcomes some of these disadvantages. 2 Vision Inspection Synthesis Our testing environment SYGEVIN (see [7]) is a knowledge-based system that automatically generates vision inspections. It accepts three types of input data: specification of vision inspection objectives, 1 Automatic Vision Programming training images, Aut,omatic design of industrial vision inspections can be viewed as a computer aided synthesis of vision programs. Systems providing this type of services are designed especially to enable efficient utilization of large vision libraries even by experts from other fields (see [I, 2, 3, 41). Comparing to general vision programs, industrial vision inspections exhibit some specialities. Used vision algorithms are relatively simple due to speed requirements. Processed images are usually of good quality because acquisition process can be controlled. On the other hand, vision inspections are executed repeatedly on many images for a long time. Image properties can vary within certain range thus the inspection has to be well adjusted. Systems generating vision inspections do not build programs from scratch. Instead, they usually compose inspection algorithms from library functions, called vision tools. To produce reliable, well adjusted inspections the systems have to utilize all available information during inspection set up. Beside training images they often retrieve a lot of context information interacting with a user. Models of inspected objects represent important pieces of such information. Systems for aided inspection design usually can accept CAD object models and set up inspections 'Address: Americka 22, 120 00 Prague, Czech Republic. E-mail: martink0rockwell.cz. pecold0rockwell.cz context information. All these pieces of information are provided by the user. Based on input data, the system generates vision inspection programs. These programs define inspection structure, i.e. sequence of selected vision tools together with data flow, and parameter values for used vision tools. The system is able to generate more inspections if its database allows alternative solutions of given task.