Using Deep Learning for the Image Recognition of Motifs on the Center of Sukhothai Ceramics

Abstract

The motifs on the center of Sukhothai ceramics are essential elements for determining the age of the ceramics. Sukhothai ceramics in each kiln were made with different pattern production techniques, and thus one specific pattern appears only in a particular kiln. Thus, archaeologists can determine which ceramic was produced from which particular kiln site by investigating its motif. Motif identification requires a well-experienced expert to identify the tracery of the pattern on the center of a ceramic. Thus, identifying such archaeological evidence is complex even for general archaeologists. The aim of this research was to study the use of deep convolutional neural networks for classifying seven motif patterns on the center of Sukhothai ceramics (i.e. Chrysanthemum bouquet, Classic scroll, Conch shell, Fish pattern, Flower head pattern, Printed Chrysanthemum head, and Tibetan Buddhist vajra ). We collected a new dataset, including 557 images of ceramics, from two kiln sites. Each ceramic’s motif was labeled by Thai ceramic experts. The collection of the motifs on the center of the Sukhothai ceramic dataset was called CMC Sukhothai Ceramic Dataset. The efficiency of the motif identification on the center of Sukhothai ceramics was evaluated by comparing five pretrained convolutional neural network models: DenseNet121, InceptionV3, VGG16, GoogLeNet, and AlexNet. Then, the models that were efficient for our dataset were selected and trained by fine tuning. Results showed that the motif recognition of VGG16 + our classification layers generated the best efficiency at 500 epochs of training and 86.54% of accuracy in the test dataset.