Using fully convolutional neural networks for object detection in robot welding applications

Abstract

In this thesis we have tried to use deep convolutional neural networks in a automated welding application. This application was the instant robot programming system from Exner. The network used was the fully convolutional U-net architecture because it performed well with few training examples, accepts inputs from most sizes and separates on a pixel level. The network was trained on a in this thesis created training set and able to separate most inputs correctly with an average recall rate of 86,80%.