Classification of fragments: recognition of artistic style

Disruptive phenomena are known for damaging the original appearance of artworks, even when the pieces are carefully and meticulously put together. Machine Learning and Computer vision techniques may be effective support tools during the works in an excavation site to resolve the issues of reconstructing frescoes, painting or sculptures. Not infrequently happens that multiple and different pictorial styles may be found in the same historical ruins. Classifying the fragments into the right pictorial style plays a significant role as a prior step for reconstruction. Determining the pictorial style or the painter from the full artwork is feasible, but it is distant from the real working conditions in the excavations. In this work a dataset called CLEOPATRA is proposed. It collects frescoes of eleven painting styles, from Prehistory to Surrealism, and simulate the effects of collapsing events that turn them into pieces. The fragmentation strategies applied are varied in terms of size (each work is divided into 10, 20, 40 and 80 pieces respectively) and the modus operandi for obtaining the pieces (one approach exploits the salient features of the work while another is random). For each of the fragments obtained, two different Machine Learning approaches were involved to solve the classification problem, discussing the advantages and critical aspect that make it a challenging but also relevant problem for archaeologists. The approach that exploits information on colour, shape and texture features is the one that gives the best results. A further study in which starting with the three most disparate artistic periods and progressively adding two more shows clearly that when the problem is significantly simpler, i.e. with fewer classes, the overall performance of the system improves.