Automated Design of Image Recognition Process for Picking System

In this study, we propose an automated design system for an image recognition algorithm applicable to picking work in general and actual factory environments. Considering that an image recognition algorithm design consists of frameworks for selecting a rough recognition method from any of the three basic procedures of pre-processing of contained images, feature-extraction, and discrimination, we formulate it as an optimization problem and propose a random multi-start optimization method by which to derive solutions. We have conducted four types of evaluation experiments for the following four combinations: large or small degrees of similarity in the shape of objects to be recognized and whether they have patterned surfaces. The evaluation experiments show that the proposed design system succeeds in selecting a framework that fits the features of the objects to be recognized and that the designed basic processes have an F measure of 0.9 or more.