A kernel support vector machine-based feature selection approach for recognizing Flying Apsaras’ streamers in the Dunhuang Grotto Murals, China

• Define the shape-based features of Flying Apsaras’ streamers.
• Propose a morphological descriptor of incorporating these features for KSVM.
• Demonstrate the suitability of the descriptor and KSVM for streamer recognition.

Recognizing Flying Apsaras’ streamers is of great importance in analyzing Chinese cultural background and art forms form the early Chinese dynasties. This analysis is very valuable for cultural protection and heritage. However, few studies have focused on recognition of Flying Apsaras in the Dunhuang Grotto Murals, China, which record elements of Chinese culture in different Chinese dynasties. By introducing a set of feature descriptors for Flying Apsaras’ streamers, this paper proposes a morphological streamer feature descriptor to describe the shape-based features (i.e., slenderness, posture ratio, area ratio, and intensity) of Flying Apsaras’ streamers. Then, a Kernel Support Vector Machine (KSVM) is implemented to locate and recognize Flying Apsaras’ streamers using the proposed feature descriptor. This machine is composed of two important parts: region segmentation of the images in the Dunhuang Grotto Murals, and KSVM-based feature selection for streamer recognition. The implemented KSVM approach incorporating the proposed morphological feature descriptor can classify streamer regions with 89.56% accuracy. Comparing the results of different classifiers and different feature descriptors demonstrates that the proposed morphological feature descriptor is a suitable morphological operator and that the KSVM is a suitable classifier for Flying Apsaras’ streamers in the Dunhuang Grotto Murals, China.