KOOPI: Keypoint-Oriented Object Positioning in Industry

Abstract

In manufacturing, object detection via industrial images enables many typical applications such as positioning of screw holes, electronic components, and other devices. The conventional method based on classical image processing generally consists of two steps: image feature extraction and object classification, so that it usually results in a low detection speed and poor accuracy due to its complicated procedure. In recent years, thanks to the rapid development of deep learning networks, higher classification accuracy and less computing performance requirement can be achieved in many typical applications, compared to using the conventional schemes. In this paper, by investigating object detection based on deep learning, a new idea utilizing some keypoint-oriented deep learning networks to the workpiece positioning area is proposed and verified by collecting dataset from practical workpieces. Our novel method performs competitively with existing schemes and runs in real-time. As for the simulation of screw hole positioning, the proposed network can effectively detect the screw hole in the image and accurately locate it. By comparing with commercial software such as Halcon® or VisionPro®, the feasibility of applying keypoint-oriented deep learning networks to intelligent manufacturing is validated.