Automatic Classification of Thangka Headdresses Based on Convolutional Depth Neural Networks

Abstract

As a representative of Tibetan culture, people's headdresses in Thangka can be divided into hairpin, monk hat and crown. In order to meet users' demand for accurate retrieval of Thangka, the category information can be used to mark headdresses of Thangka, thereby increasing the accuracy. Existing headdress classifiers suffer from a common problem: image segmentation is required before classification. When segmentation is not satisfactory, the human interaction is also required. This paper presents a classification method for Thangka headdress based on convolutional deep neural networks, without segmentation and human interaction, ease of application. First, top features of headdresses are unsupervised learned by self-encoding; then enter labeled training samples to train a softmax classifier after the convolution and pooling operation process; and finally using the test sampled to test classifier's performance. Compared with other methods, experimental results show that this method can be a good automatic classifier of headdresses, and can be more readily applied to headdress labeling.