Path Opening for Hyperspectral Crack Detection of Cultural Heritage Paintings

Crack formation in paintings is inevitable as they age, whether as a result of environment or mishandling. The ability to detect and quantify the degree to which cracks have formed would afford conservators a metric with which to assess the state of a painting and its rate of deterioration. Attempts to automate accurate crack detection in cultural heritage tend to struggle with the non-uniformity of the background, and simple morphological operators applied to grayscale images of paintings are seldom sufficient. This work investigates the extension of morphological operations (top-hat and path opening) into the hyperspectral domain to isolate long, thin structures and thereby extract a crack map from a hyperspectral image of painting. We assess the results with Intersection over Union (IoU) against a manually constructed ground truth of the crack map. We find that while IoU can give a reasonable indication of success, a visual inspection shows that workflow order has an effect on the number of false negative detections, and local contrast plays an important role in crack segmentation.