Robotic Vision Inspection of Weld Quality Using Convolutional Neural Networks

Abstract

This paper deals with ways to replace humans in visually inspecting weld quality using a robot and convolutional neural networks. Highly specialized workers are needed to inspect the quality of welds, which are currently in short supply. Therefore, the design of robotic methods of vision quality inspection is a current topic. Our approach to weld defect recognition and localization uses deep learning methods. For these purposes, a dataset of weld images was created using the robot, in which the quality inspector marked defects in the welds. In the paper, we compare methods using several architectures of convolutional neural networks.