Abstract: The objective of this study is to suggest a method for classifying archeological fragments into groups. For this task, the method suggested begins with conversion of images from their original RGB color to a Hue, Saturation and Value (HSV) color. From that point forward, a 2D median filtering algorithm is implemented to remove any resultant noise. Next, each image is separated into six sub-block of equivalent size. In order to extract the feature for each sub-block, it is represented as a vector intersection of colors between each part of the image and the corresponding parts of the five remaining images. At this stage, we obtain a vector that consists of the six values for each image. For the last stage, a Self-Organization Map (SOM) Neural Network classifies the fragments into groups relying upon their HSV color feature. The algorithm was tested on several images of pottery fragments and the results achieved demonstrate this approach is promising and is able to cluster fragments into groups with high precision.

INTRODUCTION

Technology has effectively contributed to the preservation of cultural wealth through complex automated image processing procedures and several authors have participated in providing many of the approaches for the (semi/automated) reconstruction of unknown broken or torn objects from a large number of irregular fragments (Zhu, 2013; Zhou et al., 2007), such as archaeology, forensics and medical imaging (Papaodysseus et al., 2012; Youguang et al., 2013). In particular, several researchers are interested in reassembling archaeological fragments, especially when exploring archaeological objects that have high archaeological value for the scholars such as (Papaodysseus et al., 2012; Leit?o and Stolfi, 2005).

Therefore, it is of great interest that the objects are reassembled before they are lost or damaged. Artifacts are often found in archaeological excavation sites and are randomly mixed with each other. Therefore, classifying them manually is a difficult and time consuming task, because they commonly exceed thousands of fragments. Thus, only a few previous research works focused on the classification of fragments (Makridis and Daras, 2012) such as (Papaodysseus et al., 2012; Karasik and Smilansky, 2011):

The work by Maiza and Gaildrat (2005) presented a method based on the profile of the object and classified it based on a genetic algorithm to evaluate and determine the optimum position between the fragment and the tested model. The work by Ying and Gang (2010) proposed an approach for the classification of ancient ceramic fragments by relying on surface texture features that were extracted using a Gabor wavelet transformation. The classification of ceramic was performed through applying a nonsupervision kernel- based fuzzy clustering algorithm. Also, the study of Smith et al. (2010) suggested a method depending on the color and texture features and they classified fragments on the basis of the K-Nearest Neighbor method. The work by Zhou et al. (2011) involved the classification of ancient porcelain based on features of color and texture; the fragments were classified by using the K-Nearest Neighbor method. 

Moreover, the work by Makridis and Daras (2012) focused on the automatic classification of ceramic and pottery fragments that contain little textual information. The technique is based on chromaticity and chrominance (color), the low and medium level features, as well as the K-Nearest Neighborhood classifier to classify the fragments.