Surface Defect Detection Based on Weakly Supervised Learning and Graph Convolutional Network

Abstract

The paper presents a method for detecting defects based on a convolutional network on graphs and weakly controlled learning. The defect detection algorithm includes the following steps: preliminary segmentation of the image into regions using simple linear iterative clustering method, construction of the region adjacency graph, prediction of graph nodes based on graph convolutional network on graphs. The region adjacency graph is being constructed. Each region is a node of the graph. The edges of the graph connect adjacent regions of the image. Spectral convolution of graphs is used in the proposed approach. Defects are marked on the training dataset using circles and ovals. The metric mean absolute error was used to evaluate the developed method. The method has good results when using inaccurate labeling. The structural description of the scene makes it possible to compensate for the accuracy of the annotated data to some extent.