Semantic Segmentation for Corrosion Detection in Archaeological Artefacts before Restoration

From the moment of being excavated till they become a museum exhibit, archaeological artefacts undergo a careful process of restoration, elaborately conducted by human experts with the help of complex devices. After the chemical composition of the object is approximated, the next step of the pathway is to assess the degradation of the surface, i.e. the quantification of corrosion. While earlier work proposed an automation of the step related to the estimation of the chemical concentration, the current study attempts to further offer a computational solution for the detection of corroded areas of the artefact. Iron historical items were considered, stereo microscopy images were produced and the restorers manually roughly delineated the regions containing rust. An U-Net architecture was trained on the annotated collection to recognize rust from clean areas. Even with a preliminary minimal manual delineation of the degraded zones for training, the deep learning model was able to recognize the similar areas in new objects in the test phase.