Abstract:

Objective Traditional costume culture is in danger of disappearing gradually and needs effective conservation methods. Currently, most of conservation methods rely on human resources, such as recording traditional costume by taking photos and scanning, causing that the conservation efficiency and culture exchanging are low and an efficient method to preserve the culture is lacking. Therefore, a new deep learning algorithm was proposed for highly accurate recognition of traditional costumes, and an online web identification system was designed based on cloud computing technologies. The proposed research should serve as a new alternative way for conservation and recognization of traditional costume culture efficiently.

Method Firstly, a traditional costume image dataset was constructed and enhanced by the combination of multiple background replacement and geometric transformation. Then, three modified DenseNet169 network models were built by introducing the squee and excitation (SE), convolutional block attention module (CBAM) and coordinate attention (CA), respectively, and these models were later integrated together to form a high-performance algorithm. After that, based on cloud computation and web technology, an online recognition system for typical traditional costume images was constructed by combining image normalization pre-processing and the new algorithm.

Results A traditional costume images dataset, which contains 92 160 images of a total of 15 styles, such as costumes of Zang, Man, Mongolian, Miao, Yi, Gejia, Li, Qiang, Hui, Dai, Zhuang, Han, She, Bai and Korean nationlities, was set up. The comprehensive recognition accuracies for the three improved models (using attention mechanisms SE-Dense Net 169, CBAM-DenseNet169 and CA-DenseNet169, respectively) were 89.50%, 89.83% and 90.17%, respectively (Tab.1). Although all their comprehensive recognition performances were good and similar, each model was limited by poor recognition accuracies for some specific different traditional costume categories. For example, the separated recognition accuracies of SE-DenseNet169 on Li and Zang costumes were only 77.5% and 80%, respectively. After weighting integration of the three models, the final algorithm obtained a high comprehensive recognition accuracy of 93.50% on the verification set. Compared with the previous best comprehensive recognition accuracy of CA-DenseNet169, an improvement of 3.33% was achieved. With the new algorithm, apart from relatively low separated recognition accuracy (about 87.50%) for Li costumes, the separated recognition accuracies for other traditional costume categories were all above 90.00%. Once the Korean costume image was input into the system, the most possible 3 prediction costume categories and the consumption time were displayed (Fig.4). 600 Real scene traditional costume images from different costume categories were tested, only 15 images' corrected categories were not shown in the most possible 3 prediction costume categories, which indicated a high comprehensive recognition accuracy of 98.00%. The value would be decreased to 93.50% if only using the most possible 1 prediction costume category as the output result. Meanwhile, the average processing and recognizing time taken by the system (deployed on an Aliyun server with dual-cores intel i5 CPU and 4 GiB RAM) for an image of 1 MB was around 11-13 s, which should be acceptable.

Conclusion In addition to the problems of lack of effective methods to protect traditional costume culture and limited recognization channels, the research built an online recognition system of typical traditional costume images. The system could efficiently identify 15 types of traditional costume images, and it is convenient to operate, recognize and share. Besides recording, protecting and spreading traditional costume culture efficiently, the system could also be used as a digital tool to promote tourism, culture and economy in various ethnic regions. The research would provide a new alternative solution for conserving and recognizing traditional costume culture. However, at present, the system still has some limitations, such as, only few recognizable traditional costume categories, low recognition accuracy of individual traditional costume categories and slightly slow recognition speed. In the future, the number of recognizable costume categories should be expanded, the algorithm should be improved, and the interface and operation process of the system should be optimized.

Key words: traditional costume image, dataset enhancement, transfer learning, weighted voting verdict, online recognition system