Combining image analysis and modular neural networks for classification of mineral inclusions and pores in archaeological potsherds

The approach combining image analysis techniques and artificial neural networks is proposed here for automatic classification of mineral inclusions and pores in archaeological potsherds using optical digital images. Particularly, the automatic identification of quartz, calcareous aggregates and secondary porosity is considered. A collection of both plane and cross polarised light images acquired via a digital camera connected to optical microscopy in transmitted light is used. Images concern Holocene potsherds (8900 e4200 years BP) from Takarkori rock shelter archaeological site (SW Libya, Central Sahara). The adopted methodology involves different phases. Firstly, image segmentation is carried out to isolate regions corresponding to the interested mineral inclusions and pores. A segmentation procedure based on mathematical operators is customized for each type of inclusions and for pores. Secondly, numerical features are extracted from each segmented region, thus collecting data to perform automatic classification. A modular classifier is considered for classification, which is based on a combination of three two-layer feed-forward neural networks that are trained separately to recognise each class. Experimental results show that the created modular classifier provides high classification accuracy for both inclusions and pores. The classifier was finally applied absent the image analysis phase on new samples to show the effectiveness of the proposed methodology.