Automatic Thread-Level Canvas Analysis: A machine-learning approach to analyzing the canvas of paintings

摘要

Canvas analysis is an important tool in art-historical studies, as it can provide information on whether two paintings were made on canvas that originated from the same bolt. Canvas analysis algorithms analyze radiographs of paintings to identify (ir)regularities in the spacings between the canvas threads. To reduce noise, current state-of-the-art algorithms do this by averaging the signal over a number of threads, which leads to information loss in the final measurements. This article presents an algorithm capable of performing thread-level canvas analysis: the algorithm identifies each of the individual threads in the canvas radiograph and directly measures between-distances and angles of the identified threads. We present two case studies to illustrate the potential merits of our thread-level canvas analysis algorithm, viz. on a small collection of paintings ostensibly by Nicholas Poussin and on a small collection of paintings by Vincent van Gogh.