Simultaneous design of image conversion parameters and classifier in object recognition for a picking task

In general, it is difficult to construct an object recognition system, because such a system has many design variables and often these cannot be designed independently. However, in certain manufacturing tasks, it is not always necessary to design all variables. In this study, we selected a picking task as the target task for the experiment. We restricted the design variables to parameters of the preprocessing modules and the classifier, and then proposed a method for constructing an object recognition system to achieve the intended task. We formulated the problem as an optimization problem, and proposed to solve it by using nearest neighbor search. In order to demonstrate the efficiency of our method, we conducted experiments on training images under various conditions. By using the obtained parameters and classifier, the system was able to detect all of the target objects in the target images, to recognize the objects' shapes correctly, and to estimate the objects' positions and angles within an error range that is sufficiently small for a picking task.