Craquelure as a Graph: Application of Image Processing and Graph Neural Networks to the Description of Fracture Patterns

Abstract

Cracks on a painting is not a defect but an inimitablesignature of an artwork which can be used for origin exam-ination, aging monitoring, damage identification, and evenforgery detection. This work presents the development of anew methodology and corresponding toolbox for the extrac-tion and characterization of information from an image of acraquelure pattern.The proposed approach processes craquelure network asa graph. The graph representation captures the networkstructure via mutual organization of junctions and frac-tures. Furthermore, it is invariant to any geometrical dis-tortions. At the same time, our tool extracts the propertiesof each node and edge individually, which allows to char-acterize the pattern statistically.We illustrate benefits from the graph representationand statistical features individually using novel GraphNeural Network and hand-crafted descriptors correspond-ingly. However, we also show that the best performance isachieved when both techniques are merged into one frame-work. We perform experiments on the dataset for paintingsorigin classification and demonstrate that our approachoutperforms existing techniques by a large margin.