Abstract: The design pattern of Chinese imperial costumes contains rich cultural connotation. However, due to the lack of data set of pixel-level semantic annotation, the accurate segmentation of Chinese imperial costume images has become a very challenging problem. In this paper, a bi-level model integrating deep learning and GrabCut is proposed to realize the object detection and segmentation. The characteristics of different deep convolution neural network models are analyzed, and a two-stage object detector R-FCN is selected in the object detection layer (ODL). The segmentation layer (SL) of the proposed model employs GrabCut algorithm based on graph theory to produce final segmentation result. Experiments show that the proposed bi-level model can produce good segmentation results in the Chinese imperial costume image data set.

Key words:  automatic segmentation, bi-level model, object detection layer, segmentation layer, Chinese imperial costume image