A development of easily trainable vision system for the multi-purpose dual arm robots

Small quantity batch production is a major trend in modem society and making the production line intelligent is a key for the purpose. Especially, the vision system should be able to be easily modified to recognize a variety of objects. In this paper, a novel vision system that can be trained to recognize completely new target object set of around 10 classes in an hour is proposed. An hour limitation is a design purpose that is derived from the consideration that the training should be so simple and fast to be applied on the site. The training process is performed simply by repeatedly showing the target objects to the system. Also the proposed system provides online evaluation functionality that can confirm the performance of the trained classifier before the application to the line. Through the experiments on two target object sets, the proposed vision system is found that it could be trained to recognize the new object sets in 30 minutes including the training and the evaluation achieving the high accuracy of more than 98% in normal office environment.