Specification of Image Features for the Sensory Inspection by Limit Samples

Abstract

This paper proposes a new framework for automating sensory inspections in which contents are described by limit samples alone. A bottom-up approach is presently attempting to cope with such complex inspections. Identification of the psychological structure of a human inspector is studied by analyzing the data of evaluation experiments, and the relation between human impressions and image features is examined in order to automate the sensory inspection.
In this paper, the appearance inspection of an automobile windshield by a projected checker pattern is treated as an example. Sensory scores are calculated by the comparative evaluation of defective CG images generated by a defect simulator. The relation between the sensory scores of inspectors and image features by image processing is learned by neural networks. The image features that contribute to the human judgment of the inspection are specified by a sensitivity analysis of the neural networks.