Part inspection by developmental vision.

Online part inspection is integral to modern machining and manufacturing processes. It includes dimensional verification and surface inspection which consists of the detection and classification of surface flaws. A special challenge is introduced by a reconfigurable manufacturing system (RMS), where product changes may frequently occur in the same production line. Quick adaptation of the inspection system to these changes is essential for fast ramp-up and cost effectiveness of the reconfigurable manufacturing system. Manual inspection is labor intensive and subject to human-dependent variables and errors. By contrast, machine vision inspection systems improve inspection speed and quality and reduce labor and cost. Nevertheless, an implementable artificial vision model for general and unknown environment is still illusive. As a result, existing vision systems are not capable of automatically addressing large manufacturing variations and part changes. In this dissertation we present a novel, biologically-inspired, adaptive inspection method that addresses these needs. It utilizes a developmental vision approach for automatic adaptation to training experience. It requires two processes, a developmental process and a performance process, occurring concurrently, online, in real time. The method addresses both attention and recognition problems in an unknown inspection environment. We introduce the concept of late-invariance which is achieved by a sufficiently complete set of representations. We demonstrate the adaptation of the method to different tasks by testing it on various applications: textural surface inspection, dimensional landmark detection, object setting recognition, and porosity detection. The results, showing also fast recognition and low error rate, confirm the potential industrial real-time online compatibility of this approach. The contribution of the presented inspection approach is in ramp-up time reduction in RMS, improved quality control in variable environment, and reduced inspection costs due to reduction in inspector and vision developer manpower. Specifically, we developed guidelines to the implementation of adaptive online part inspection by developmental vision, with detailed examples. This included definitions of various inspection tasks as visual recognition problems and their solutions; directions for creating invariance to a general property in industrial conditions; and instructions for matching acquisition and illumination systems with the underlined approach.Ph.D.Applied SciencesArtificial intelligenceComputer scienceMechanical engineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/124006/2/3121881.pd