TY - CPAPER
KW - Contrast Learning
KW - Contrast learning
KW - Deep learning
KW - Deep learning
KW - Digital protection
KW - Few-Shot Learning
KW - Few-shot learning
KW - Image coding
KW - Intangible cultural heritages
KW - Modern technologies
KW - Object detection
KW - Object recognition
KW - Objects detection
KW - target detection
KW - Targets detection
KW - Thangka Image
KW - Thangka image
KW - Tibetans
AU - H. Tang
AU - C. Yue
AU - W. Hu
AU - L. Qiao
AB - As one of China s intangible cultural heritage, Thangka is the main carrier of Tibetan culture. The application of modern technology in the digital protection of thangka images is of great significance to promote cultural exchange and inheritance. In this paper, the headdress, seat and handheld objects of the thangka images are taken as the detection targets, and they are classified and recognized by using the method,which is few-shot object detection by contrastive proposal coding. This method introduces contrastive learning into the few-shot object detection method. That takes Faster R-CNN as the baseline model, adds a contrastive branch parallel to the classification and prediction branch at the end of the model. In the contrastive branch, the problem of misclassification is alleviated by calculating the similarity score between the object proposal boxes. The experiment shows that this method can achieve good results on the thangka data set, with mAP reaching 38.1\% and AP50 reaching 51.3\%.
C2 - Int. Conf. Intell. Comput. Signal Process., ICSP
DO - 10.1109/ICSP54964.2022.9778717
N1 - Journal Abbreviation: Int. Conf. Intell. Comput. Signal Process., ICSP
N2 - As one of China s intangible cultural heritage, Thangka is the main carrier of Tibetan culture. The application of modern technology in the digital protection of thangka images is of great significance to promote cultural exchange and inheritance. In this paper, the headdress, seat and handheld objects of the thangka images are taken as the detection targets, and they are classified and recognized by using the method,which is few-shot object detection by contrastive proposal coding. This method introduces contrastive learning into the few-shot object detection method. That takes Faster R-CNN as the baseline model, adds a contrastive branch parallel to the classification and prediction branch at the end of the model. In the contrastive branch, the problem of misclassification is alleviated by calculating the similarity score between the object proposal boxes. The experiment shows that this method can achieve good results on the thangka data set, with mAP reaching 38.1\% and AP50 reaching 51.3\%.
PB - Institute of Electrical and Electronics Engineers Inc.
SN - 9781665478571 (ISBN)
SP - 1916
EP - 1919
TI - Object Detection of Few-Shot Thangka Images by Contrastive Proposal Coding
UR - https://www.scopus.com/inward/record.uri?eid=2-s2.0-85131818997&doi=10.1109%2fICSP54964.2022.9778717&partnerID=40&md5=0fa29c1adac1759b4a3ccb0333847bd5
ER -