Machine Part Recognition based on Transfer Learning and Knowledge Graph

Automatic classification of parts based on computer vision allows to reduce the costs of such labor-intensive operations as searching, tracking, sorting of parts. The method of recognition of machine parts based on transfer learning and a knowledge graph is presented in the paper. A knowledge graph is built based on images from the source and target areas. Images from the target area are obtained under real observation conditions and may differ slightly from the base class. There are significantly fewer images from the target area than images from the source area. Graph convolutional network is used to predict the labels of graph nodes. Thus, additional information based on the knowledge graph is used to increase the accuracy of recognition of machine parts. The method allows to increase the precision and recall values by 1–6 %. The method reduces the dependence of machine vision systems on the preliminary training stage and allows to determine preferences for specific systems.