Automatic control of class weights in the semantic segmentation of corrosion compounds on archaeological artefacts

Abstract

The semantic segmentation for irregularly and not uniformly disposed patterns becomes even more difficult when the occurrence of categories is imbalanced within the images. One example is represented by heavily corroded artefacts in archaeological digs. The current study therefore proposes a weighted loss function within a deep learning architecture for semantic segmentation of corrosion compounds from microscopy images of archaeological objects, where the values for the class weights are generated via genetic algorithms. The fitness evaluation of individuals is the estimation that a surrogate of the deep learner gives concerning the segmentation accuracy. The obtained class weight values are compared to a random search through the space of potential configurations and another automated means to compute them, in terms of resulting model accuracy.