From image to graph: Topological representation of Wire harnesses using deep learning

Abstract

The ability to classify an abstract representation into an image of a wire harness is a key step in automating the wire harness assembly industries. This study examines different methods and machine learning architectures to extract visual and physical features from an image of a wire harness. To identify suitable methods for abstractly representing and classifying a wire harness, a systematic literature review was con ducted. The systematic literature review was not limited to studies about wire harnesses, but expanded to studies about deformable linear objects. Following the systematic literature review, an experimental study was conducted that evaluated an existing implementation for graph representation of a wire harness and a novel method based on a You Only Live Once (YOLO) segmentation model together with a Graph Convolutional Network (GCN) to classify the graph representation against a validation file of known harnesses. The segmentation output from the YOLO network is fed into a skeletonization method that returns a graph that represents the wire harness. In order to classify the graph representation against known harness structures, a GCN along with a validation step is performed. The models are trained on a proprietary dataset and two open-source datasets. The proposed solution can be easily adapted to classify new wire harnesses.